Upload 75 files

.gitattributes

@@ -33,3 +33,22 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+examples/instrumental/blue_instrumental.wav filter=lfs diff=lfs merge=lfs -text
+examples/instrumental/Counter_clockwise_Clock_instrumental.wav filter=lfs diff=lfs merge=lfs -text
+examples/instrumental/one_last_kiss_instrumental.wav filter=lfs diff=lfs merge=lfs -text
+examples/song/blue.wav filter=lfs diff=lfs merge=lfs -text
+examples/song/Counter_clockwise_Clock.wav filter=lfs diff=lfs merge=lfs -text
+examples/song/one_last_kiss.wav filter=lfs diff=lfs merge=lfs -text
+examples/vocals/blue_vocal.wav filter=lfs diff=lfs merge=lfs -text
+examples/vocals/Counter_clockwise_Clock_vocal.wav filter=lfs diff=lfs merge=lfs -text
+examples/vocals/one_last_kiss_vocal.wav filter=lfs diff=lfs merge=lfs -text
+output_2stems/blue-instrumental.wav filter=lfs diff=lfs merge=lfs -text
+output_2stems/blue-vocals.wav filter=lfs diff=lfs merge=lfs -text
+output_2stems/temp-instrumental.wav filter=lfs diff=lfs merge=lfs -text
+output_2stems/temp-vocals.wav filter=lfs diff=lfs merge=lfs -text
+raw/1.wav filter=lfs diff=lfs merge=lfs -text
+spleeter/2stems_instrumental.pdparams filter=lfs diff=lfs merge=lfs -text
+spleeter/2stems_vocals.pdparams filter=lfs diff=lfs merge=lfs -text
+trained_models/纳西妲.pdparams filter=lfs diff=lfs merge=lfs -text
+trained_models/派蒙.pdparams filter=lfs diff=lfs merge=lfs -text
+trained_models/YH.pdparams filter=lfs diff=lfs merge=lfs -text

.ipynb_checkpoints/build.gradio-checkpoint.py

@@ -0,0 +1,176 @@
+import io
+import os
+
+import gradio as gr
+import librosa
+import numpy as np
+import soundfile
+from inference.infer_tool import Svc
+import logging
+import os
+import paddle
+import requests
+import utils
+from spleeter import Separator
+
+build_dir=os.getcwd()
+if build_dir == "/home/aistudio":
+    build_dir += "/build"
+
+model_dir=build_dir+'/trained_models'
+
+model_list_path = model_dir + "/model_list.txt"
+
+# 筛选出文件夹
+models = []
+for filename in os.listdir(model_dir):
+    # 判断文件名是否以 '.pdparams' 结尾，并且不包含后缀部分
+    if filename.endswith('.pdparams') and os.path.splitext(filename)[0].isalpha():
+        models.append(os.path.splitext(filename)[0])
+cache_model = {}
+
+def reboot():  
+    os.execv(sys.executable, ['python'] + sys.argv)
+
+def separate_fn(song_input):
+    try:
+        if song_input is None:
+            return "请上传歌曲",None,None,None,None
+        params_2stems = {
+        'sample_rate': 44100,
+        'frame_length': 4096,
+        'frame_step': 1024,
+        'T': 512,
+        'F': 1024,
+        'num_instruments': ['vocals', 'instrumental'],
+        'output_dir': build_dir+'/output_2stems',
+        'checkpoint_path': build_dir+'/spleeter',
+        'use_elu': False}
+        sampling_rate, song = song_input
+        soundfile.write("temp.wav", song, sampling_rate, format="wav")
+        # 初始化分离器
+        sep = Separator(params_2stems)
+        sep.separate('temp.wav')
+        vocal_path = params_2stems["output_dir"]+"/temp-vocals.wav"
+        instrumental_path = params_2stems["output_dir"]+"/temp-instrumental.wav"
+        return "分离成功，请继续前往体验【转换】和【混音】",vocal_path,instrumental_path,vocal_path,instrumental_path
+    except Exception as e:
+        import traceback
+        return traceback.format_exc() , None,None,None,None
+
+
+def convert_fn(model_name, input_audio,input_audio_micro, vc_transform, auto_f0,cluster_ratio, slice_db, noise_scale):
+    try:
+        if model_name in cache_model:
+            model = cache_model[model_name]
+        else:
+            if paddle.device.is_compiled_with_cuda()==False and len(cache_model)!=0:
+                return f"目前运行环境为CPU，受制于平台算力，每次启动本项目只允许加载1个模型，当前已加载{next(iter(cache_model))}",None,None
+            config_path = f"{build_dir}/trained_models/config.json"
+            model = Svc(f"{build_dir}/trained_models/{model_name}.pdparams", config_path,mode="test")
+            cache_model[model_name] = model
+        if input_audio is None and input_audio_micro is None:
+            return "请上传音频", None,None
+        if input_audio_micro is not None:
+            input_audio = input_audio_micro
+        sampling_rate, audio = input_audio
+        duration = audio.shape[0] / sampling_rate
+        audio = (audio / np.iinfo(audio.dtype).max).astype(np.float32)
+        if len(audio.shape) > 1:
+            audio = librosa.to_mono(audio.transpose(1, 0))
+        if sampling_rate != 16000:
+            audio = librosa.resample(audio, orig_sr=sampling_rate, target_sr=16000)
+        print(audio.shape)
+        out_wav_path = "temp.wav"
+        soundfile.write(out_wav_path, audio, 16000, format="wav")
+        print(cluster_ratio, auto_f0, noise_scale)
+        _audio = model.slice_inference(out_wav_path, model_name, vc_transform, slice_db, cluster_ratio, auto_f0, noise_scale)
+        del model
+        return "转换成功，请继续前往体验【混音】", (44100, _audio),(44100, _audio)
+    except Exception as e:
+        import traceback
+        return traceback.format_exc() , None,None
+
+def compose_fn(input_vocal,input_instrumental,mixing_ratio=0.5):
+    try:
+        outlog = "混音成功"
+        if input_vocal is None:
+            return "请上传人声", None
+        if input_instrumental is None:
+            return "请上传伴奏", None
+        vocal_sampling_rate, vocal = input_vocal
+        vocal_duration = vocal.shape[0] / vocal_sampling_rate
+        vocal = (vocal / np.iinfo(vocal.dtype).max).astype(np.float32)
+        if len(vocal.shape) > 1:
+            vocal = librosa.to_mono(vocal.transpose(1, 0))
+        if vocal_sampling_rate != 44100:
+            vocal = librosa.resample(vocal, orig_sr=vocal_sampling_rate, target_sr=44100)
+
+        instrumental_sampling_rate, instrumental = input_instrumental
+        instrumental_duration = instrumental.shape[0] / instrumental_sampling_rate
+        instrumental = (instrumental / np.iinfo(instrumental.dtype).max).astype(np.float32)
+        if len(instrumental.shape) > 1:
+            instrumental = librosa.to_mono(instrumental.transpose(1, 0))
+        if instrumental_sampling_rate != 44100:
+            instrumental = librosa.resample(instrumental, orig_sr=instrumental_sampling_rate, target_sr=44100)
+        if len(vocal)!=len(instrumental):
+            min_length = min(len(vocal),len(instrumental))
+            instrumental = instrumental[:min_length]
+            vocal = vocal[:min_length]
+            outlog = "人声伴奏长度不一致，已自动截断较长的音频"
+
+        mixed_audio = (1 - mixing_ratio) * vocal + mixing_ratio * instrumental
+        mixed_audio_data = mixed_audio.astype(np.float32)
+        return outlog,(44100,mixed_audio_data)
+    except Exception as e:
+        import traceback
+        return traceback.format_exc() , None
+
+
+app = gr.Blocks()
+with app:
+    gr.Markdown('<h1 style="text-align: center;">SVC歌声转换全流程体验（伴奏分离，转换，混音）</h1>')
+    btn_reboot = gr.Button("重启程序", variant="primary")
+    with gr.Tabs() as tabs:
+        with gr.TabItem("人声伴奏分离"):
+            gr.Markdown('<p>该项目人声分离的效果弱于UVR5，如自备分离好的伴奏和人声可跳过该步骤</p>')
+            song_input = gr.Audio(label="上传歌曲（tips:上传后点击右上角✏可以进行歌曲剪辑）",interactive=True)
+            gr.Examples(examples=[build_dir+"/examples/song/blue.wav",build_dir+"/examples/song/Counter_clockwise_Clock.wav",build_dir+"/examples/song/one_last_kiss.wav"],inputs=song_input,label="歌曲样例")
+
+            btn_separate = gr.Button("人声伴奏分离", variant="primary")
+            text_output1 = gr.Textbox(label="输出信息")
+            vocal_output1 = gr.Audio(label="输出人声",interactive=False)
+            instrumental_output1 = gr.Audio(label="输出伴奏",interactive=False)
+        with gr.TabItem("转换"):
+            model_name = gr.Dropdown(label="模型", choices=models, value="纳西妲")
+            vocal_input1 = gr.Audio(label="上传人声",interactive=True)
+            gr.Examples(examples=[build_dir+"/examples/vocals/blue_vocal.wav",build_dir+"/examples/vocals/Counter_clockwise_Clock_vocal.wav",build_dir+"/examples/vocals/one_last_kiss_vocal.wav"],inputs=vocal_input1,label="人声样例")
+            btn_use_separate = gr.Button("使用【人声伴奏分离】分离的人声")
+            micro_input = gr.Audio(label="麦克风输入（优先于上传的人声）",source="microphone",interactive=True)
+            vc_transform = gr.Number(label="变调（半音数量,升八度12降八度-12）", value=0)
+            cluster_ratio = gr.Number(label="聚类模型混合比例", value=0,visible=False)
+            auto_f0 = gr.Checkbox(label="自动预测音高（转换歌声时不要打开，会严重跑调）", value=False)
+            slice_db = gr.Number(label="静音分贝阈值（嘈杂的音频可以-30，干声保留呼吸可以-50）", value=-50)
+            noise_scale = gr.Number(label="noise_scale", value=0.2)
+            btn_convert = gr.Button("转换", variant="primary")
+            text_output2 = gr.Textbox(label="输出信息")
+            vc_output2 = gr.Audio(label="输出音频",interactive=False)
+
+        with gr.TabItem("混音"):
+            vocal_input2 = gr.Audio(label="上传人声",interactive=True)
+            btn_use_convert = gr.Button("使用【转换】输出的人声")
+            instrumental_input1 = gr.Audio(label="上传伴奏")
+            gr.Examples(examples=[build_dir+"/examples/instrumental/blue_instrumental.wav",build_dir+"/examples/instrumental/Counter_clockwise_Clock_instrumental.wav",build_dir+"/examples/instrumental/one_last_kiss_instrumental.wav"],inputs=instrumental_input1,label="伴奏样例")
+            btn_use_separate2 = gr.Button("使用【人声伴奏分离】分离的伴奏")
+            mixing_ratio = gr.Slider(0, 1, value=0.75,step=0.01,label="混音比例（人声:伴奏）", info="人声:伴奏")
+            btn_compose = gr.Button("混音", variant="primary")
+            text_output3 = gr.Textbox(label="输出信息")
+            song_output = gr.Audio(label="输出歌曲",interactive=False)
+        btn_separate.click(separate_fn, song_input, [text_output1, vocal_output1,instrumental_output1,vocal_input1,instrumental_input1])
+        btn_convert.click(convert_fn, [model_name, vocal_input1,micro_input,vc_transform,auto_f0,cluster_ratio, slice_db, noise_scale], [text_output2, vc_output2,vocal_input2])
+        btn_reboot.click(reboot)
+        btn_use_convert.click(lambda x:x,vc_output2,vocal_input2)
+        btn_use_separate.click(lambda x:x,vocal_output1,vocal_input1)
+        btn_use_separate2.click(lambda x:x,instrumental_output1,instrumental_input1)
+
+app.launch()

build.gradio.py

@@ -0,0 +1,176 @@
+import io
+import os
+
+import gradio as gr
+import librosa
+import numpy as np
+import soundfile
+from inference.infer_tool import Svc
+import logging
+import os
+import paddle
+import requests
+import utils
+from spleeter import Separator
+
+build_dir=os.getcwd()
+if build_dir == "/home/aistudio":
+    build_dir += "/build"
+
+model_dir=build_dir+'/trained_models'
+
+model_list_path = model_dir + "/model_list.txt"
+
+# 筛选出文件夹
+models = []
+for filename in os.listdir(model_dir):
+    # 判断文件名是否以 '.pdparams' 结尾，并且不包含后缀部分
+    if filename.endswith('.pdparams') and os.path.splitext(filename)[0].isalpha():
+        models.append(os.path.splitext(filename)[0])
+cache_model = {}
+
+def reboot():   
+    os._exit(0)
+
+def separate_fn(song_input):
+    try:
+        if song_input is None:
+            return "请上传歌曲",None,None,None,None
+        params_2stems = {
+        'sample_rate': 44100,
+        'frame_length': 4096,
+        'frame_step': 1024,
+        'T': 512,
+        'F': 1024,
+        'num_instruments': ['vocals', 'instrumental'],
+        'output_dir': build_dir+'/output_2stems',
+        'checkpoint_path': build_dir+'/spleeter',
+        'use_elu': False}
+        sampling_rate, song = song_input
+        soundfile.write("temp.wav", song, sampling_rate, format="wav")
+        # 初始化分离器
+        sep = Separator(params_2stems)
+        sep.separate('temp.wav')
+        vocal_path = params_2stems["output_dir"]+"/temp-vocals.wav"
+        instrumental_path = params_2stems["output_dir"]+"/temp-instrumental.wav"
+        return "分离成功，请继续前往体验【转换】和【混音】",vocal_path,instrumental_path,vocal_path,instrumental_path
+    except Exception as e:
+        import traceback
+        return traceback.format_exc() , None,None,None,None
+
+
+def convert_fn(model_name, input_audio,input_audio_micro, vc_transform, auto_f0,cluster_ratio, slice_db, noise_scale):
+    try:
+        if model_name in cache_model:
+            model = cache_model[model_name]
+        else:
+            if paddle.device.is_compiled_with_cuda()==False and len(cache_model)!=0:
+                return f"目前运行环境为CPU，受制于平台算力，每次启动本项目只允许加载1个模型，当前已加载{next(iter(cache_model))}",None,None
+            config_path = f"{build_dir}/trained_models/config.json"
+            model = Svc(f"{build_dir}/trained_models/{model_name}.pdparams", config_path,mode="test")
+            cache_model[model_name] = model
+        if input_audio is None and input_audio_micro is None:
+            return "请上传音频", None,None
+        if input_audio_micro is not None:
+            input_audio = input_audio_micro
+        sampling_rate, audio = input_audio
+        duration = audio.shape[0] / sampling_rate
+        audio = (audio / np.iinfo(audio.dtype).max).astype(np.float32)
+        if len(audio.shape) > 1:
+            audio = librosa.to_mono(audio.transpose(1, 0))
+        if sampling_rate != 16000:
+            audio = librosa.resample(audio, orig_sr=sampling_rate, target_sr=16000)
+        print(audio.shape)
+        out_wav_path = "temp.wav"
+        soundfile.write(out_wav_path, audio, 16000, format="wav")
+        print(cluster_ratio, auto_f0, noise_scale)
+        _audio = model.slice_inference(out_wav_path, model_name, vc_transform, slice_db, cluster_ratio, auto_f0, noise_scale)
+        del model
+        return "转换成功，请继续前往体验【混音】", (44100, _audio),(44100, _audio)
+    except Exception as e:
+        import traceback
+        return traceback.format_exc() , None,None
+
+def compose_fn(input_vocal,input_instrumental,mixing_ratio=0.5):
+    try:
+        outlog = "混音成功"
+        if input_vocal is None:
+            return "请上传人声", None
+        if input_instrumental is None:
+            return "请上传伴奏", None
+        vocal_sampling_rate, vocal = input_vocal
+        vocal_duration = vocal.shape[0] / vocal_sampling_rate
+        vocal = (vocal / np.iinfo(vocal.dtype).max).astype(np.float32)
+        if len(vocal.shape) > 1:
+            vocal = librosa.to_mono(vocal.transpose(1, 0))
+        if vocal_sampling_rate != 44100:
+            vocal = librosa.resample(vocal, orig_sr=vocal_sampling_rate, target_sr=44100)
+
+        instrumental_sampling_rate, instrumental = input_instrumental
+        instrumental_duration = instrumental.shape[0] / instrumental_sampling_rate
+        instrumental = (instrumental / np.iinfo(instrumental.dtype).max).astype(np.float32)
+        if len(instrumental.shape) > 1:
+            instrumental = librosa.to_mono(instrumental.transpose(1, 0))
+        if instrumental_sampling_rate != 44100:
+            instrumental = librosa.resample(instrumental, orig_sr=instrumental_sampling_rate, target_sr=44100)
+        if len(vocal)!=len(instrumental):
+            min_length = min(len(vocal),len(instrumental))
+            instrumental = instrumental[:min_length]
+            vocal = vocal[:min_length]
+            outlog = "人声伴奏长度不一致，已自动截断较长的音频"
+
+        mixed_audio = (1 - mixing_ratio) * vocal + mixing_ratio * instrumental
+        mixed_audio_data = mixed_audio.astype(np.float32)
+        return outlog,(44100,mixed_audio_data)
+    except Exception as e:
+        import traceback
+        return traceback.format_exc() , None
+
+
+app = gr.Blocks()
+with app:
+    gr.Markdown('<h1 style="text-align: center;">SVC歌声转换全流程体验（伴奏分离，转换，混音）</h1>')
+    btn_reboot = gr.Button("重启程序", variant="primary")
+    with gr.Tabs() as tabs:
+        with gr.TabItem("人声伴奏分离"):
+            gr.Markdown('<p>该项目人声分离的效果弱于UVR5，如自备分离好的伴奏和人声可跳过该步骤</p>')
+            song_input = gr.Audio(label="上传歌曲（tips:上传后点击右上角✏可以进行歌曲剪辑）",interactive=True)
+            gr.Examples(examples=[build_dir+"/examples/song/blue.wav",build_dir+"/examples/song/Counter_clockwise_Clock.wav",build_dir+"/examples/song/one_last_kiss.wav"],inputs=song_input,label="歌曲样例")
+
+            btn_separate = gr.Button("人声伴奏分离", variant="primary")
+            text_output1 = gr.Textbox(label="输出信息")
+            vocal_output1 = gr.Audio(label="输出人声",interactive=False)
+            instrumental_output1 = gr.Audio(label="输出伴奏",interactive=False)
+        with gr.TabItem("转换"):
+            model_name = gr.Dropdown(label="模型", choices=models, value="纳西妲")
+            vocal_input1 = gr.Audio(label="上传人声",interactive=True)
+            gr.Examples(examples=[build_dir+"/examples/vocals/blue_vocal.wav",build_dir+"/examples/vocals/Counter_clockwise_Clock_vocal.wav",build_dir+"/examples/vocals/one_last_kiss_vocal.wav"],inputs=vocal_input1,label="人声样例")
+            btn_use_separate = gr.Button("使用【人声伴奏分离】分离的人声")
+            micro_input = gr.Audio(label="麦克风输入（优先于上传的人声）",source="microphone",interactive=True)
+            vc_transform = gr.Number(label="变调（半音数量,升八度12降八度-12）", value=0)
+            cluster_ratio = gr.Number(label="聚类模型混合比例", value=0,visible=False)
+            auto_f0 = gr.Checkbox(label="自动预测音高（转换歌声时不要打开，会严重跑调）", value=False)
+            slice_db = gr.Number(label="静音分贝阈值（嘈杂的音频可以-30，干声保留呼吸可以-50）", value=-50)
+            noise_scale = gr.Number(label="noise_scale", value=0.2)
+            btn_convert = gr.Button("转换", variant="primary")
+            text_output2 = gr.Textbox(label="输出信息")
+            vc_output2 = gr.Audio(label="输出音频",interactive=False)
+
+        with gr.TabItem("混音"):
+            vocal_input2 = gr.Audio(label="上传人声",interactive=True)
+            btn_use_convert = gr.Button("使用【转换】输出的人声")
+            instrumental_input1 = gr.Audio(label="上传伴奏")
+            gr.Examples(examples=[build_dir+"/examples/instrumental/blue_instrumental.wav",build_dir+"/examples/instrumental/Counter_clockwise_Clock_instrumental.wav",build_dir+"/examples/instrumental/one_last_kiss_instrumental.wav"],inputs=instrumental_input1,label="伴奏样例")
+            btn_use_separate2 = gr.Button("使用【人声伴奏分离】分离的伴奏")
+            mixing_ratio = gr.Slider(0, 1, value=0.75,step=0.01,label="混音比例（人声:伴奏）", info="人声:伴奏")
+            btn_compose = gr.Button("混音", variant="primary")
+            text_output3 = gr.Textbox(label="输出信息")
+            song_output = gr.Audio(label="输出歌曲",interactive=False)
+        btn_separate.click(separate_fn, song_input, [text_output1, vocal_output1,instrumental_output1,vocal_input1,instrumental_input1])
+        btn_convert.click(convert_fn, [model_name, vocal_input1,micro_input,vc_transform,auto_f0,cluster_ratio, slice_db, noise_scale], [text_output2, vc_output2,vocal_input2])
+        btn_reboot.click(reboot)
+        btn_use_convert.click(lambda x:x,vc_output2,vocal_input2)
+        btn_use_separate.click(lambda x:x,vocal_output1,vocal_input1)
+        btn_use_separate2.click(lambda x:x,instrumental_output1,instrumental_input1)
+
+app.launch()

cluster/__init__.py

@@ -0,0 +1,29 @@
+import numpy as np
+import paddle
+from sklearn.cluster import KMeans
+
+def get_cluster_model(ckpt_path:str):
+    checkpoint = paddle.load(ckpt_path)
+    kmeans_dict = {}
+    for spk, ckpt in checkpoint.items():
+        km = KMeans(ckpt["n_features_in_"])
+        km.__dict__["n_features_in_"] = ckpt["n_features_in_"]
+        km.__dict__["_n_threads"] = ckpt["_n_threads"]
+        km.__dict__["cluster_centers_"] = ckpt["cluster_centers_"]
+        kmeans_dict[spk] = km
+    return kmeans_dict
+
+def get_cluster_result(model, x, speaker):
+    """
+        x: np.array [t, 256]
+        return cluster class result
+    """
+    return model[speaker].predict(x)
+
+def get_cluster_center_result(model, x,speaker):
+    """x: np.array [t, 256]"""
+    predict = model[speaker].predict(x)
+    return model[speaker].cluster_centers_[predict]
+
+def get_center(model, x,speaker):
+    return model[speaker].cluster_centers_[x]

cluster/train_cluster.py

@@ -0,0 +1,88 @@
+import os
+from glob import glob
+from pathlib import Path
+import paddle
+import logging
+import argparse
+import numpy as np
+from sklearn.cluster import KMeans, MiniBatchKMeans
+import tqdm
+logging.basicConfig(level=logging.INFO)
+logger = logging.getLogger(__name__)
+import time
+import random
+
+def train_cluster(in_dir, n_clusters, use_minibatch=True, verbose=False):
+
+    logger.info(f"正在从{in_dir}加载特征")
+    features = []
+    nums = 0
+    for path in tqdm.tqdm(in_dir.glob("*.soft.pdtensor")):
+        path = str(path)
+        features.append(paddle.load(path).squeeze(0).numpy().T)
+        # print(features[-1].shape)
+    features = np.concatenate(features, axis=0)
+    print(nums, features.nbytes/ 1024**2, "MB , 形状：",features.shape, features.dtype)
+    features = features.astype(np.float32)
+    logger.info(f"聚类特征的形状：{features.shape}")
+    t = time.time()
+    if use_minibatch:
+        kmeans = MiniBatchKMeans(n_clusters=n_clusters,verbose=verbose, batch_size=4096, max_iter=80).fit(features)
+    else:
+        kmeans = KMeans(n_clusters=n_clusters,verbose=verbose).fit(features)
+    print(time.time()-t, "s")
+
+    x = {
+            "n_features_in_": kmeans.n_features_in_,
+            "_n_threads": kmeans._n_threads,
+            "cluster_centers_": kmeans.cluster_centers_,
+    }
+    print("结束")
+
+    return x
+
+
+if __name__ == "__main__":
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--dataset', type=Path, default="./dataset/44k",
+                        help='path of training data directory')
+    parser.add_argument('--output', type=Path, default="logs/44k",
+                        help='path of model output directory')
+
+    args = parser.parse_args()
+
+    checkpoint_dir = args.output
+    dataset = args.dataset
+    n_clusters = 10000
+
+    ckpt = {}
+    for spk in os.listdir(dataset):
+        if os.path.isdir(dataset/spk):
+            print(f"正在给{spk}训练kmeans中……")
+            in_dir = dataset/spk
+            x = train_cluster(in_dir, n_clusters, verbose=False)
+            ckpt[spk] = x
+
+    checkpoint_path = checkpoint_dir / f"kmeans_{n_clusters}.pdparams"
+    checkpoint_path.parent.mkdir(exist_ok=True, parents=True)
+    paddle.save(
+        ckpt,
+        str(checkpoint_path),
+    )
+
+
+    # import cluster
+    # for spk in tqdm.tqdm(os.listdir("dataset")):
+    #     if os.path.isdir(f"dataset/{spk}"):
+    #         print(f"start kmeans inference for {spk}...")
+    #         for feature_path in tqdm.tqdm(glob(f"dataset/{spk}/*.discrete.npy", recursive=True)):
+    #             mel_path = feature_path.replace(".discrete.npy",".mel.npy")
+    #             mel_spectrogram = np.load(mel_path)
+    #             feature_len = mel_spectrogram.shape[-1]
+    #             c = np.load(feature_path)
+    #             c = utils.tools.repeat_expand_2d(torch.FloatTensor(c), feature_len).numpy()
+    #             feature = c.T
+    #             feature_class = cluster.get_cluster_result(feature, spk)
+    #             np.save(feature_path.replace(".discrete.npy", ".discrete_class.npy"), feature_class)
+

configs/config.json

@@ -0,0 +1,95 @@
+{
+    "train": {
+        "log_interval": 800,
+        "eval_interval": 400,
+        "seed": 1234,
+        "epochs": 114514,
+        "learning_rate": 0.0001,
+        "betas": [
+            0.8,
+            0.99
+        ],
+        "eps": 1e-05,
+        "batch_size": 2,
+        "fp16_run": true,
+        "lr_decay": 0.999875,
+        "segment_size": 10240,
+        "init_lr_ratio": 1,
+        "warmup_epochs": 0,
+        "c_mel": 45,
+        "c_kl": 1.0,
+        "use_sr": true,
+        "max_speclen": 512,
+        "port": "8001",
+        "keep_ckpts": 5
+    },
+    "data": {
+        "training_files": "filelists/train.txt",
+        "validation_files": "filelists/val.txt",
+        "max_wav_value": 32768.0,
+        "sampling_rate": 44100,
+        "filter_length": 2048,
+        "hop_length": 512,
+        "win_length": 2048,
+        "n_mel_channels": 80,
+        "mel_fmin": 0.0,
+        "mel_fmax": 22050
+    },
+    "model": {
+        "inter_channels": 192,
+        "hidden_channels": 192,
+        "filter_channels": 768,
+        "n_heads": 2,
+        "n_layers": 6,
+        "kernel_size": 3,
+        "p_dropout": 0.1,
+        "resblock": "1",
+        "resblock_kernel_sizes": [
+            3,
+            7,
+            11
+        ],
+        "resblock_dilation_sizes": [
+            [
+                1,
+                3,
+                5
+            ],
+            [
+                1,
+                3,
+                5
+            ],
+            [
+                1,
+                3,
+                5
+            ]
+        ],
+        "upsample_rates": [
+            8,
+            8,
+            2,
+            2,
+            2
+        ],
+        "upsample_initial_channel": 512,
+        "upsample_kernel_sizes": [
+            16,
+            16,
+            4,
+            4,
+            4
+        ],
+        "n_layers_q": 3,
+        "use_spectral_norm": false,
+        "gin_channels": 256,
+        "ssl_dim": 256,
+        "n_speakers": 200
+    },
+    "spk": {
+        "yuuka": 0
+    },
+    "clean_logs": true,
+    "trainer": "admin"
+}

data_utils.py

@@ -0,0 +1,143 @@
+import time
+import os
+import random
+import numpy as np
+import paddle
+
+import modules.commons as commons
+import utils
+from modules.mel_processing import spectrogram_torch, spec_to_mel_torch
+from utils import load_wav_to_torch, load_filepaths_and_text
+
+# import h5py
+
+
+"""Multi speaker version"""
+
+
+class TextAudioSpeakerLoader(paddle.io.Dataset):
+    """
+        1) loads audio, speaker_id, text pairs
+        2) normalizes text and converts them to sequences of integers
+        3) computes spectrograms from audio files.
+    """
+
+    def __init__(self, audiopaths, hparams):
+        self.audiopaths = load_filepaths_and_text(audiopaths)
+        self.max_wav_value = hparams.data.max_wav_value
+        self.sampling_rate = hparams.data.sampling_rate
+        self.filter_length = hparams.data.filter_length
+        self.hop_length = hparams.data.hop_length
+        self.win_length = hparams.data.win_length
+        self.sampling_rate = hparams.data.sampling_rate
+        self.use_sr = hparams.train.use_sr
+        self.spec_len = hparams.train.max_speclen
+        self.spk_map = hparams.spk
+
+        random.seed(1234)
+        random.shuffle(self.audiopaths)
+
+    def get_audio(self, filename):
+        filename = filename.replace("\\", "/")
+        audio, sampling_rate = load_wav_to_torch(filename)
+        if sampling_rate != self.sampling_rate:
+            raise ValueError("{} SR doesn't match target {} SR".format(
+                sampling_rate, self.sampling_rate))
+        audio_norm = audio / self.max_wav_value
+        audio_norm = audio_norm.unsqueeze(0)
+        spec_filename = filename.replace(".wav", ".spec.pdtensor")
+        if os.path.exists(spec_filename):
+            spec = paddle.load(spec_filename)
+        else:
+            spec = spectrogram_torch(audio_norm, self.filter_length,
+                                     self.sampling_rate, self.hop_length, self.win_length,
+                                     center=False)
+            spec = paddle.squeeze(spec, 0)
+            paddle.save(spec, spec_filename)
+
+        spk = filename.split("/")[-2]
+        spk = paddle.to_tensor([self.spk_map[spk]],dtype = 'int64')
+
+        f0 = np.load(filename + ".f0.npy")
+        f0, uv = utils.interpolate_f0(f0)
+        f0 = paddle.to_tensor(f0,dtype = 'float32')
+        uv = paddle.to_tensor(uv,dtype = 'float32')
+
+        c = paddle.load(filename+ ".soft.pdtensor")
+        c = utils.repeat_expand_2d(c.squeeze(0), f0.shape[0])
+
+
+        lmin = min(c.shape[-1], spec.shape[-1])
+        assert abs(c.shape[-1] - spec.shape[-1]) < 3, (c.shape[-1], spec.shape[-1], f0.shape, filename)
+        assert abs(audio_norm.shape[1]-lmin * self.hop_length) < 3 * self.hop_length
+        spec, c, f0, uv = spec[:, :lmin], c[:, :lmin], f0[:lmin], uv[:lmin]
+        audio_norm = audio_norm[:, :lmin * self.hop_length]
+        # if spec.shape[1] < 30:
+        #     print("skip too short audio:", filename)
+        #     return None
+        if spec.shape[1] > 800:
+            start = random.randint(0, spec.shape[1]-800)
+            end = start + 790
+            spec, c, f0, uv = spec[:, start:end], c[:, start:end], f0[start:end], uv[start:end]
+            audio_norm = audio_norm[:, start * self.hop_length : end * self.hop_length]
+
+        return c, f0, spec, audio_norm, spk, uv
+
+    def __getitem__(self, index):
+        return self.get_audio(self.audiopaths[index][0])
+
+    def __len__(self):
+        return len(self.audiopaths)
+
+
+class TextAudioCollate:
+
+    def __call__(self, batch):
+        batch = [b for b in batch if b is not None]
+
+        input_lengths, ids_sorted_decreasing = \
+        paddle.sort(
+            paddle.to_tensor([x[0].shape[1] for x in batch],dtype = 'int64'),
+            axis=0, descending=True),\
+        paddle.argsort(
+            paddle.to_tensor([x[0].shape[1] for x in batch],dtype = 'int64'),
+            axis=0, descending=True)
+
+        max_c_len = max([x[0].shape[1] for x in batch])
+        max_wav_len = max([x[3].shape[1] for x in batch])
+
+        lengths = paddle.zeros((len(batch),),dtype = 'int64')
+        c_padded = paddle.to_tensor(np.random.rand(len(batch), batch[0][0].shape[0], max_c_len),dtype = 'float32')
+        f0_padded =  paddle.to_tensor(np.random.rand(len(batch), max_c_len),dtype = 'float32')
+        spec_padded =  paddle.to_tensor(np.random.rand(len(batch), batch[0][2].shape[0], max_c_len),dtype = 'float32')
+        wav_padded =  paddle.to_tensor(np.random.rand(len(batch), 1, max_wav_len),dtype = 'float32')
+        spkids = paddle.zeros((len(batch), 1),dtype = 'int64')
+        uv_padded = paddle.to_tensor(np.random.rand(len(batch), max_c_len),dtype = 'float32')
+
+        c_padded.zero_()
+        spec_padded.zero_()
+        f0_padded.zero_()
+        wav_padded.zero_()
+        uv_padded.zero_()
+
+        for i in range(len(ids_sorted_decreasing)):
+            row = batch[ids_sorted_decreasing[i]]
+
+            c = row[0]
+            c_padded[i, :, :c.shape[1]] = c
+            lengths[i] = c.shape[1]
+
+            f0 = row[1]
+            f0_padded[i, :f0.shape[0]] = f0
+
+            spec = row[2]
+            spec_padded[i, :, :spec.shape[1]] = spec
+
+            wav = row[3]
+            wav_padded[i, :, :wav.shape[1]] = wav
+            spkids[i, 0] = row[4]
+
+            uv = row[5]
+            uv_padded[i, :uv.shape[0]] = uv
+
+        return c_padded, f0_padded, spec_padded, wav_padded, spkids, lengths, uv_padded

examples/instrumental/Counter_clockwise_Clock_instrumental.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4a1c1d882e28458a0aa448cb68b65344e4839a2354766d679f20970c30362fc9
+size 2062380

examples/instrumental/blue_instrumental.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7bbbc0b86619c634cbad2769ce035e96d4b7e598d50260be11a84e21f7b8b311
+size 5420756

examples/instrumental/one_last_kiss_instrumental.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:223b2eb0ef567c891bf61a3f05d97aea97b18fb44bc21d6726053f81c5fe0f57
+size 6575226

examples/song/Counter_clockwise_Clock.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b87df2edb949b8c8d87f740ebd701c764425636906b0232c6ea7375b0dc898e8
+size 2252296

examples/song/blue.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2b80bd01d14fcd86bbee2c048f94b8b5927ae0a661ff5b2b99a014110a8c9a49
+size 2709736

examples/song/one_last_kiss.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fc671b5095cc1dfa00772de09f8fa94741482b573eafbd88188fe909cdb1f173
+size 6575226

examples/vocals/Counter_clockwise_Clock_vocal.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f749f2b9a1c8af5afbc66544956337ce12fd1c5b9d64e69d1212847a36480b9e
+size 2062380

examples/vocals/blue_vocal.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5014db90a4bd53a18054ae2a2f9b1e733c8c474a6872beace1c3bd716b2cb61f
+size 2484268

examples/vocals/one_last_kiss_vocal.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d52245bb0ac3965ebedcf730d5e21bc8d35519b84184c40b9f8ec780420b813c
+size 6568602

filelists/test.txt

@@ -0,0 +1,2 @@
+./dataset/44k/yuuka/7_93.wav
+./dataset/44k/yuuka/1_98.wav

filelists/train.txt

@@ -0,0 +1,857 @@
+./dataset/44k/yuuka/1_44.wav
+./dataset/44k/yuuka/6_3.wav
+./dataset/44k/yuuka/3_136.wav
+./dataset/44k/yuuka/3_63.wav
+./dataset/44k/yuuka/4_14.wav
+./dataset/44k/yuuka/7_168.wav
+./dataset/44k/yuuka/4_12.wav
+./dataset/44k/yuuka/364983.wav
+./dataset/44k/yuuka/606448.wav
+./dataset/44k/yuuka/5_21.wav
+./dataset/44k/yuuka/8_23.wav
+./dataset/44k/yuuka/9_69.wav
+./dataset/44k/yuuka/5_91.wav
+./dataset/44k/yuuka/3_163.wav
+./dataset/44k/yuuka/6_27.wav
+./dataset/44k/yuuka/234705.wav
+./dataset/44k/yuuka/3_176.wav
+./dataset/44k/yuuka/1_93.wav
+./dataset/44k/yuuka/76490.wav
+./dataset/44k/yuuka/3_179.wav
+./dataset/44k/yuuka/1_69.wav
+./dataset/44k/yuuka/192554.wav
+./dataset/44k/yuuka/218550.wav
+./dataset/44k/yuuka/6_16.wav
+./dataset/44k/yuuka/5_35.wav
+./dataset/44k/yuuka/207846.wav
+./dataset/44k/yuuka/7_145.wav
+./dataset/44k/yuuka/audio_28.wav
+./dataset/44k/yuuka/2_30.wav
+./dataset/44k/yuuka/1_55.wav
+./dataset/44k/yuuka/5_6.wav
+./dataset/44k/yuuka/3_178.wav
+./dataset/44k/yuuka/3_19.wav
+./dataset/44k/yuuka/audio_15.wav
+./dataset/44k/yuuka/3_134.wav
+./dataset/44k/yuuka/1_1.wav
+./dataset/44k/yuuka/5_66.wav
+./dataset/44k/yuuka/7_61.wav
+./dataset/44k/yuuka/7_43.wav
+./dataset/44k/yuuka/9_4.wav
+./dataset/44k/yuuka/4_42.wav
+./dataset/44k/yuuka/7_130.wav
+./dataset/44k/yuuka/5_52.wav
+./dataset/44k/yuuka/919255.wav
+./dataset/44k/yuuka/4_82.wav
+./dataset/44k/yuuka/1_114.wav
+./dataset/44k/yuuka/7_7.wav
+./dataset/44k/yuuka/1_79.wav
+./dataset/44k/yuuka/1_33.wav
+./dataset/44k/yuuka/9_21.wav
+./dataset/44k/yuuka/574054.wav
+./dataset/44k/yuuka/69301.wav
+./dataset/44k/yuuka/5_30.wav
+./dataset/44k/yuuka/119619.wav
+./dataset/44k/yuuka/7_167.wav
+./dataset/44k/yuuka/1_45.wav
+./dataset/44k/yuuka/6_23.wav
+./dataset/44k/yuuka/656221.wav
+./dataset/44k/yuuka/9_65.wav
+./dataset/44k/yuuka/3_85.wav
+./dataset/44k/yuuka/7_188.wav
+./dataset/44k/yuuka/6_13.wav
+./dataset/44k/yuuka/3_60.wav
+./dataset/44k/yuuka/997011.wav
+./dataset/44k/yuuka/audio_1.wav
+./dataset/44k/yuuka/6_39.wav
+./dataset/44k/yuuka/5_15.wav
+./dataset/44k/yuuka/7775.wav
+./dataset/44k/yuuka/969593.wav
+./dataset/44k/yuuka/7_94.wav
+./dataset/44k/yuuka/39442.wav
+./dataset/44k/yuuka/9_37.wav
+./dataset/44k/yuuka/3_46.wav
+./dataset/44k/yuuka/5_41.wav
+./dataset/44k/yuuka/517653.wav
+./dataset/44k/yuuka/9_38.wav
+./dataset/44k/yuuka/2_18.wav
+./dataset/44k/yuuka/255268.wav
+./dataset/44k/yuuka/8_5.wav
+./dataset/44k/yuuka/4_22.wav
+./dataset/44k/yuuka/1_81.wav
+./dataset/44k/yuuka/3_39.wav
+./dataset/44k/yuuka/262686.wav
+./dataset/44k/yuuka/3_105.wav
+./dataset/44k/yuuka/571657.wav
+./dataset/44k/yuuka/8_20.wav
+./dataset/44k/yuuka/6_33.wav
+./dataset/44k/yuuka/3_165.wav
+./dataset/44k/yuuka/594447.wav
+./dataset/44k/yuuka/7_181.wav
+./dataset/44k/yuuka/695541.wav
+./dataset/44k/yuuka/1_32.wav
+./dataset/44k/yuuka/1_92.wav
+./dataset/44k/yuuka/3_40.wav
+./dataset/44k/yuuka/98954.wav
+./dataset/44k/yuuka/8_11.wav
+./dataset/44k/yuuka/7_25.wav
+./dataset/44k/yuuka/4_43.wav
+./dataset/44k/yuuka/3_26.wav
+./dataset/44k/yuuka/7_23.wav
+./dataset/44k/yuuka/1_34.wav
+./dataset/44k/yuuka/3_24.wav
+./dataset/44k/yuuka/3_135.wav
+./dataset/44k/yuuka/4_33.wav
+./dataset/44k/yuuka/9_39.wav
+./dataset/44k/yuuka/52347.wav
+./dataset/44k/yuuka/3_54.wav
+./dataset/44k/yuuka/3_145.wav
+./dataset/44k/yuuka/1_13.wav
+./dataset/44k/yuuka/463536.wav
+./dataset/44k/yuuka/3_104.wav
+./dataset/44k/yuuka/2_20.wav
+./dataset/44k/yuuka/3_66.wav
+./dataset/44k/yuuka/7_182.wav
+./dataset/44k/yuuka/3_1.wav
+./dataset/44k/yuuka/5_63.wav
+./dataset/44k/yuuka/3_12.wav
+./dataset/44k/yuuka/audio_40.wav
+./dataset/44k/yuuka/7_127.wav
+./dataset/44k/yuuka/1_11.wav
+./dataset/44k/yuuka/audio_10.wav
+./dataset/44k/yuuka/490130.wav
+./dataset/44k/yuuka/3_162.wav
+./dataset/44k/yuuka/audio_11.wav
+./dataset/44k/yuuka/3_42.wav
+./dataset/44k/yuuka/635392.wav
+./dataset/44k/yuuka/7_52.wav
+./dataset/44k/yuuka/5_51.wav
+./dataset/44k/yuuka/8_6.wav
+./dataset/44k/yuuka/8_7.wav
+./dataset/44k/yuuka/8_1.wav
+./dataset/44k/yuuka/3_94.wav
+./dataset/44k/yuuka/1_10.wav
+./dataset/44k/yuuka/9_76.wav
+./dataset/44k/yuuka/7_160.wav
+./dataset/44k/yuuka/3_98.wav
+./dataset/44k/yuuka/553418.wav
+./dataset/44k/yuuka/3_53.wav
+./dataset/44k/yuuka/295642.wav
+./dataset/44k/yuuka/5_45.wav
+./dataset/44k/yuuka/5_50.wav
+./dataset/44k/yuuka/1_47.wav
+./dataset/44k/yuuka/4_50.wav
+./dataset/44k/yuuka/609212.wav
+./dataset/44k/yuuka/4_77.wav
+./dataset/44k/yuuka/5_13.wav
+./dataset/44k/yuuka/2_21.wav
+./dataset/44k/yuuka/9_25.wav
+./dataset/44k/yuuka/2_6.wav
+./dataset/44k/yuuka/4_1.wav
+./dataset/44k/yuuka/8_4.wav
+./dataset/44k/yuuka/1_14.wav
+./dataset/44k/yuuka/7_55.wav
+./dataset/44k/yuuka/3_90.wav
+./dataset/44k/yuuka/7_98.wav
+./dataset/44k/yuuka/1_26.wav
+./dataset/44k/yuuka/7_187.wav
+./dataset/44k/yuuka/622956.wav
+./dataset/44k/yuuka/4_37.wav
+./dataset/44k/yuuka/61332.wav
+./dataset/44k/yuuka/1_43.wav
+./dataset/44k/yuuka/233242.wav
+./dataset/44k/yuuka/audio_36.wav
+./dataset/44k/yuuka/2_3.wav
+./dataset/44k/yuuka/9_80.wav
+./dataset/44k/yuuka/3_14.wav
+./dataset/44k/yuuka/9_32.wav
+./dataset/44k/yuuka/audio_12.wav
+./dataset/44k/yuuka/7_164.wav
+./dataset/44k/yuuka/4_56.wav
+./dataset/44k/yuuka/5_36.wav
+./dataset/44k/yuuka/359799.wav
+./dataset/44k/yuuka/6_0.wav
+./dataset/44k/yuuka/5_85.wav
+./dataset/44k/yuuka/6_4.wav
+./dataset/44k/yuuka/537888.wav
+./dataset/44k/yuuka/7_108.wav
+./dataset/44k/yuuka/352307.wav
+./dataset/44k/yuuka/3_18.wav
+./dataset/44k/yuuka/3_51.wav
+./dataset/44k/yuuka/8_26.wav
+./dataset/44k/yuuka/7_49.wav
+./dataset/44k/yuuka/3_129.wav
+./dataset/44k/yuuka/2_13.wav
+./dataset/44k/yuuka/7_44.wav
+./dataset/44k/yuuka/3_50.wav
+./dataset/44k/yuuka/5_95.wav
+./dataset/44k/yuuka/audio_29.wav
+./dataset/44k/yuuka/6_21.wav
+./dataset/44k/yuuka/8_15.wav
+./dataset/44k/yuuka/963257.wav
+./dataset/44k/yuuka/7_95.wav
+./dataset/44k/yuuka/2_11.wav
+./dataset/44k/yuuka/4_38.wav
+./dataset/44k/yuuka/1_68.wav
+./dataset/44k/yuuka/1_58.wav
+./dataset/44k/yuuka/3_181.wav
+./dataset/44k/yuuka/3_29.wav
+./dataset/44k/yuuka/7_76.wav
+./dataset/44k/yuuka/2_24.wav
+./dataset/44k/yuuka/6_7.wav
+./dataset/44k/yuuka/7_134.wav
+./dataset/44k/yuuka/3_89.wav
+./dataset/44k/yuuka/audio_19.wav
+./dataset/44k/yuuka/3_41.wav
+./dataset/44k/yuuka/9_59.wav
+./dataset/44k/yuuka/7_56.wav
+./dataset/44k/yuuka/3_167.wav
+./dataset/44k/yuuka/1_31.wav
+./dataset/44k/yuuka/8_24.wav
+./dataset/44k/yuuka/4_54.wav
+./dataset/44k/yuuka/7_53.wav
+./dataset/44k/yuuka/7_120.wav
+./dataset/44k/yuuka/3_132.wav
+./dataset/44k/yuuka/9_11.wav
+./dataset/44k/yuuka/3_65.wav
+./dataset/44k/yuuka/5_8.wav
+./dataset/44k/yuuka/3_2.wav
+./dataset/44k/yuuka/519313.wav
+./dataset/44k/yuuka/audio_16.wav
+./dataset/44k/yuuka/283655.wav
+./dataset/44k/yuuka/6_41.wav
+./dataset/44k/yuuka/5_74.wav
+./dataset/44k/yuuka/5_88.wav
+./dataset/44k/yuuka/9_56.wav
+./dataset/44k/yuuka/2_23.wav
+./dataset/44k/yuuka/3_22.wav
+./dataset/44k/yuuka/4_41.wav
+./dataset/44k/yuuka/9_47.wav
+./dataset/44k/yuuka/4_5.wav
+./dataset/44k/yuuka/974105.wav
+./dataset/44k/yuuka/2_2.wav
+./dataset/44k/yuuka/86716.wav
+./dataset/44k/yuuka/300145.wav
+./dataset/44k/yuuka/9_60.wav
+./dataset/44k/yuuka/3_36.wav
+./dataset/44k/yuuka/3_77.wav
+./dataset/44k/yuuka/287748.wav
+./dataset/44k/yuuka/7_37.wav
+./dataset/44k/yuuka/3_11.wav
+./dataset/44k/yuuka/5_3.wav
+./dataset/44k/yuuka/1_62.wav
+./dataset/44k/yuuka/8_3.wav
+./dataset/44k/yuuka/audio_5.wav
+./dataset/44k/yuuka/9_14.wav
+./dataset/44k/yuuka/5_25.wav
+./dataset/44k/yuuka/5_24.wav
+./dataset/44k/yuuka/1_66.wav
+./dataset/44k/yuuka/484944.wav
+./dataset/44k/yuuka/171951.wav
+./dataset/44k/yuuka/102566.wav
+./dataset/44k/yuuka/3_27.wav
+./dataset/44k/yuuka/3_108.wav
+./dataset/44k/yuuka/audio_9.wav
+./dataset/44k/yuuka/4_80.wav
+./dataset/44k/yuuka/3_56.wav
+./dataset/44k/yuuka/9_63.wav
+./dataset/44k/yuuka/7_51.wav
+./dataset/44k/yuuka/1_51.wav
+./dataset/44k/yuuka/5_90.wav
+./dataset/44k/yuuka/356679.wav
+./dataset/44k/yuuka/724651.wav
+./dataset/44k/yuuka/7_65.wav
+./dataset/44k/yuuka/251091.wav
+./dataset/44k/yuuka/9_10.wav
+./dataset/44k/yuuka/7_2.wav
+./dataset/44k/yuuka/4_10.wav
+./dataset/44k/yuuka/3_182.wav
+./dataset/44k/yuuka/3_55.wav
+./dataset/44k/yuuka/1_9.wav
+./dataset/44k/yuuka/3_97.wav
+./dataset/44k/yuuka/1_25.wav
+./dataset/44k/yuuka/96411.wav
+./dataset/44k/yuuka/7_31.wav
+./dataset/44k/yuuka/427153.wav
+./dataset/44k/yuuka/4_48.wav
+./dataset/44k/yuuka/1_89.wav
+./dataset/44k/yuuka/audio_30.wav
+./dataset/44k/yuuka/135370.wav
+./dataset/44k/yuuka/7_166.wav
+./dataset/44k/yuuka/7_39.wav
+./dataset/44k/yuuka/4_39.wav
+./dataset/44k/yuuka/6_36.wav
+./dataset/44k/yuuka/8_9.wav
+./dataset/44k/yuuka/4_24.wav
+./dataset/44k/yuuka/4_72.wav
+./dataset/44k/yuuka/5_58.wav
+./dataset/44k/yuuka/7_162.wav
+./dataset/44k/yuuka/6_11.wav
+./dataset/44k/yuuka/audio_13.wav
+./dataset/44k/yuuka/3_67.wav
+./dataset/44k/yuuka/9_15.wav
+./dataset/44k/yuuka/3_93.wav
+./dataset/44k/yuuka/8_25.wav
+./dataset/44k/yuuka/1_74.wav
+./dataset/44k/yuuka/7_137.wav
+./dataset/44k/yuuka/9_74.wav
+./dataset/44k/yuuka/3_131.wav
+./dataset/44k/yuuka/3_31.wav
+./dataset/44k/yuuka/3_95.wav
+./dataset/44k/yuuka/445323.wav
+./dataset/44k/yuuka/1_48.wav
+./dataset/44k/yuuka/audio_23.wav
+./dataset/44k/yuuka/5_62.wav
+./dataset/44k/yuuka/5_67.wav
+./dataset/44k/yuuka/4_74.wav
+./dataset/44k/yuuka/audio_35.wav
+./dataset/44k/yuuka/9_3.wav
+./dataset/44k/yuuka/379745.wav
+./dataset/44k/yuuka/6_17.wav
+./dataset/44k/yuuka/audio_4.wav
+./dataset/44k/yuuka/3_160.wav
+./dataset/44k/yuuka/9_5.wav
+./dataset/44k/yuuka/9_22.wav
+./dataset/44k/yuuka/1_4.wav
+./dataset/44k/yuuka/2_1.wav
+./dataset/44k/yuuka/9_51.wav
+./dataset/44k/yuuka/6_38.wav
+./dataset/44k/yuuka/3_30.wav
+./dataset/44k/yuuka/3_180.wav
+./dataset/44k/yuuka/5_69.wav
+./dataset/44k/yuuka/3_44.wav
+./dataset/44k/yuuka/3_45.wav
+./dataset/44k/yuuka/7_17.wav
+./dataset/44k/yuuka/audio_41.wav
+./dataset/44k/yuuka/1_24.wav
+./dataset/44k/yuuka/7_132.wav
+./dataset/44k/yuuka/9_44.wav
+./dataset/44k/yuuka/9_43.wav
+./dataset/44k/yuuka/1_15.wav
+./dataset/44k/yuuka/469368.wav
+./dataset/44k/yuuka/144449.wav
+./dataset/44k/yuuka/2_7.wav
+./dataset/44k/yuuka/audio_17.wav
+./dataset/44k/yuuka/5_92.wav
+./dataset/44k/yuuka/7_96.wav
+./dataset/44k/yuuka/5_53.wav
+./dataset/44k/yuuka/6_20.wav
+./dataset/44k/yuuka/8_19.wav
+./dataset/44k/yuuka/5_12.wav
+./dataset/44k/yuuka/3_38.wav
+./dataset/44k/yuuka/7_57.wav
+./dataset/44k/yuuka/6_32.wav
+./dataset/44k/yuuka/3_48.wav
+./dataset/44k/yuuka/5_0.wav
+./dataset/44k/yuuka/7_1.wav
+./dataset/44k/yuuka/3_52.wav
+./dataset/44k/yuuka/7_135.wav
+./dataset/44k/yuuka/1_20.wav
+./dataset/44k/yuuka/206478.wav
+./dataset/44k/yuuka/7_6.wav
+./dataset/44k/yuuka/7_165.wav
+./dataset/44k/yuuka/271096.wav
+./dataset/44k/yuuka/1_17.wav
+./dataset/44k/yuuka/6_9.wav
+./dataset/44k/yuuka/3_122.wav
+./dataset/44k/yuuka/7_10.wav
+./dataset/44k/yuuka/audio_38.wav
+./dataset/44k/yuuka/507694.wav
+./dataset/44k/yuuka/1_28.wav
+./dataset/44k/yuuka/9_7.wav
+./dataset/44k/yuuka/8_27.wav
+./dataset/44k/yuuka/7_8.wav
+./dataset/44k/yuuka/8_29.wav
+./dataset/44k/yuuka/2_12.wav
+./dataset/44k/yuuka/8_12.wav
+./dataset/44k/yuuka/7_26.wav
+./dataset/44k/yuuka/4_9.wav
+./dataset/44k/yuuka/3_21.wav
+./dataset/44k/yuuka/3_61.wav
+./dataset/44k/yuuka/64918.wav
+./dataset/44k/yuuka/7_172.wav
+./dataset/44k/yuuka/5_27.wav
+./dataset/44k/yuuka/256460.wav
+./dataset/44k/yuuka/1_7.wav
+./dataset/44k/yuuka/7_185.wav
+./dataset/44k/yuuka/3_62.wav
+./dataset/44k/yuuka/5_40.wav
+./dataset/44k/yuuka/6_30.wav
+./dataset/44k/yuuka/7_32.wav
+./dataset/44k/yuuka/4_76.wav
+./dataset/44k/yuuka/3_185.wav
+./dataset/44k/yuuka/622091.wav
+./dataset/44k/yuuka/3_33.wav
+./dataset/44k/yuuka/9_70.wav
+./dataset/44k/yuuka/9_66.wav
+./dataset/44k/yuuka/7_40.wav
+./dataset/44k/yuuka/813875.wav
+./dataset/44k/yuuka/7_176.wav
+./dataset/44k/yuuka/4_55.wav
+./dataset/44k/yuuka/3_91.wav
+./dataset/44k/yuuka/4_4.wav
+./dataset/44k/yuuka/5_72.wav
+./dataset/44k/yuuka/551231.wav
+./dataset/44k/yuuka/7_64.wav
+./dataset/44k/yuuka/7_143.wav
+./dataset/44k/yuuka/1_37.wav
+./dataset/44k/yuuka/5_54.wav
+./dataset/44k/yuuka/5_26.wav
+./dataset/44k/yuuka/audio_0.wav
+./dataset/44k/yuuka/3_125.wav
+./dataset/44k/yuuka/1_41.wav
+./dataset/44k/yuuka/7_142.wav
+./dataset/44k/yuuka/4_23.wav
+./dataset/44k/yuuka/3_184.wav
+./dataset/44k/yuuka/6_6.wav
+./dataset/44k/yuuka/9_73.wav
+./dataset/44k/yuuka/3_166.wav
+./dataset/44k/yuuka/2_28.wav
+./dataset/44k/yuuka/9_0.wav
+./dataset/44k/yuuka/1_12.wav
+./dataset/44k/yuuka/6_18.wav
+./dataset/44k/yuuka/349028.wav
+./dataset/44k/yuuka/547091.wav
+./dataset/44k/yuuka/audio_14.wav
+./dataset/44k/yuuka/5_34.wav
+./dataset/44k/yuuka/4_19.wav
+./dataset/44k/yuuka/7_27.wav
+./dataset/44k/yuuka/7_20.wav
+./dataset/44k/yuuka/9_35.wav
+./dataset/44k/yuuka/8_16.wav
+./dataset/44k/yuuka/7_102.wav
+./dataset/44k/yuuka/3_6.wav
+./dataset/44k/yuuka/798678.wav
+./dataset/44k/yuuka/915260.wav
+./dataset/44k/yuuka/7_103.wav
+./dataset/44k/yuuka/5_18.wav
+./dataset/44k/yuuka/3_13.wav
+./dataset/44k/yuuka/7_85.wav
+./dataset/44k/yuuka/85191.wav
+./dataset/44k/yuuka/5_76.wav
+./dataset/44k/yuuka/3_4.wav
+./dataset/44k/yuuka/3_142.wav
+./dataset/44k/yuuka/4_59.wav
+./dataset/44k/yuuka/711779.wav
+./dataset/44k/yuuka/7_104.wav
+./dataset/44k/yuuka/7_129.wav
+./dataset/44k/yuuka/470902.wav
+./dataset/44k/yuuka/1_49.wav
+./dataset/44k/yuuka/7_11.wav
+./dataset/44k/yuuka/6_1.wav
+./dataset/44k/yuuka/1_67.wav
+./dataset/44k/yuuka/7_170.wav
+./dataset/44k/yuuka/1_39.wav
+./dataset/44k/yuuka/550982.wav
+./dataset/44k/yuuka/7_125.wav
+./dataset/44k/yuuka/9_75.wav
+./dataset/44k/yuuka/8_21.wav
+./dataset/44k/yuuka/3_0.wav
+./dataset/44k/yuuka/1_18.wav
+./dataset/44k/yuuka/audio_42.wav
+./dataset/44k/yuuka/5_20.wav
+./dataset/44k/yuuka/4_46.wav
+./dataset/44k/yuuka/253501.wav
+./dataset/44k/yuuka/7_184.wav
+./dataset/44k/yuuka/8_30.wav
+./dataset/44k/yuuka/2_0.wav
+./dataset/44k/yuuka/7_140.wav
+./dataset/44k/yuuka/9_77.wav
+./dataset/44k/yuuka/7_89.wav
+./dataset/44k/yuuka/56932.wav
+./dataset/44k/yuuka/audio_32.wav
+./dataset/44k/yuuka/7_29.wav
+./dataset/44k/yuuka/7_50.wav
+./dataset/44k/yuuka/9_64.wav
+./dataset/44k/yuuka/3_34.wav
+./dataset/44k/yuuka/4_36.wav
+./dataset/44k/yuuka/1_38.wav
+./dataset/44k/yuuka/9_34.wav
+./dataset/44k/yuuka/705069.wav
+./dataset/44k/yuuka/7_54.wav
+./dataset/44k/yuuka/3_121.wav
+./dataset/44k/yuuka/7_177.wav
+./dataset/44k/yuuka/audio_3.wav
+./dataset/44k/yuuka/437916.wav
+./dataset/44k/yuuka/7_169.wav
+./dataset/44k/yuuka/6_24.wav
+./dataset/44k/yuuka/2_5.wav
+./dataset/44k/yuuka/5_48.wav
+./dataset/44k/yuuka/3_47.wav
+./dataset/44k/yuuka/3_78.wav
+./dataset/44k/yuuka/3_187.wav
+./dataset/44k/yuuka/7_124.wav
+./dataset/44k/yuuka/3_169.wav
+./dataset/44k/yuuka/321860.wav
+./dataset/44k/yuuka/5_46.wav
+./dataset/44k/yuuka/3_86.wav
+./dataset/44k/yuuka/3_87.wav
+./dataset/44k/yuuka/1_83.wav
+./dataset/44k/yuuka/1_36.wav
+./dataset/44k/yuuka/1_116.wav
+./dataset/44k/yuuka/687395.wav
+./dataset/44k/yuuka/1_35.wav
+./dataset/44k/yuuka/7_186.wav
+./dataset/44k/yuuka/9_49.wav
+./dataset/44k/yuuka/8_18.wav
+./dataset/44k/yuuka/6_14.wav
+./dataset/44k/yuuka/5_4.wav
+./dataset/44k/yuuka/9_58.wav
+./dataset/44k/yuuka/498387.wav
+./dataset/44k/yuuka/5_42.wav
+./dataset/44k/yuuka/3_17.wav
+./dataset/44k/yuuka/4_3.wav
+./dataset/44k/yuuka/7_24.wav
+./dataset/44k/yuuka/314228.wav
+./dataset/44k/yuuka/652599.wav
+./dataset/44k/yuuka/7_28.wav
+./dataset/44k/yuuka/3_140.wav
+./dataset/44k/yuuka/7_3.wav
+./dataset/44k/yuuka/915926.wav
+./dataset/44k/yuuka/3_183.wav
+./dataset/44k/yuuka/7_13.wav
+./dataset/44k/yuuka/3_137.wav
+./dataset/44k/yuuka/968783.wav
+./dataset/44k/yuuka/1_52.wav
+./dataset/44k/yuuka/4_66.wav
+./dataset/44k/yuuka/5_2.wav
+./dataset/44k/yuuka/5_9.wav
+./dataset/44k/yuuka/5_70.wav
+./dataset/44k/yuuka/7_83.wav
+./dataset/44k/yuuka/8_17.wav
+./dataset/44k/yuuka/5_7.wav
+./dataset/44k/yuuka/1_70.wav
+./dataset/44k/yuuka/5_89.wav
+./dataset/44k/yuuka/3_9.wav
+./dataset/44k/yuuka/3_84.wav
+./dataset/44k/yuuka/1_29.wav
+./dataset/44k/yuuka/6_15.wav
+./dataset/44k/yuuka/884738.wav
+./dataset/44k/yuuka/1_72.wav
+./dataset/44k/yuuka/168875.wav
+./dataset/44k/yuuka/833141.wav
+./dataset/44k/yuuka/7_68.wav
+./dataset/44k/yuuka/7_41.wav
+./dataset/44k/yuuka/4_0.wav
+./dataset/44k/yuuka/7_48.wav
+./dataset/44k/yuuka/9_81.wav
+./dataset/44k/yuuka/4_44.wav
+./dataset/44k/yuuka/958019.wav
+./dataset/44k/yuuka/9_50.wav
+./dataset/44k/yuuka/5_68.wav
+./dataset/44k/yuuka/3_32.wav
+./dataset/44k/yuuka/7_106.wav
+./dataset/44k/yuuka/1_16.wav
+./dataset/44k/yuuka/5_16.wav
+./dataset/44k/yuuka/3_20.wav
+./dataset/44k/yuuka/502529.wav
+./dataset/44k/yuuka/237547.wav
+./dataset/44k/yuuka/3_186.wav
+./dataset/44k/yuuka/audio_27.wav
+./dataset/44k/yuuka/5_5.wav
+./dataset/44k/yuuka/7_35.wav
+./dataset/44k/yuuka/3_120.wav
+./dataset/44k/yuuka/5_59.wav
+./dataset/44k/yuuka/7_78.wav
+./dataset/44k/yuuka/650180.wav
+./dataset/44k/yuuka/audio_24.wav
+./dataset/44k/yuuka/678092.wav
+./dataset/44k/yuuka/6_2.wav
+./dataset/44k/yuuka/5_23.wav
+./dataset/44k/yuuka/1_88.wav
+./dataset/44k/yuuka/240071.wav
+./dataset/44k/yuuka/1_46.wav
+./dataset/44k/yuuka/15940.wav
+./dataset/44k/yuuka/8_22.wav
+./dataset/44k/yuuka/4_2.wav
+./dataset/44k/yuuka/5_93.wav
+./dataset/44k/yuuka/7_87.wav
+./dataset/44k/yuuka/454791.wav
+./dataset/44k/yuuka/5_73.wav
+./dataset/44k/yuuka/1_0.wav
+./dataset/44k/yuuka/7_4.wav
+./dataset/44k/yuuka/9_1.wav
+./dataset/44k/yuuka/586033.wav
+./dataset/44k/yuuka/1_23.wav
+./dataset/44k/yuuka/2_27.wav
+./dataset/44k/yuuka/5_38.wav
+./dataset/44k/yuuka/4_34.wav
+./dataset/44k/yuuka/347125.wav
+./dataset/44k/yuuka/2_15.wav
+./dataset/44k/yuuka/6_12.wav
+./dataset/44k/yuuka/1_84.wav
+./dataset/44k/yuuka/975179.wav
+./dataset/44k/yuuka/2_25.wav
+./dataset/44k/yuuka/4_65.wav
+./dataset/44k/yuuka/9_72.wav
+./dataset/44k/yuuka/6_8.wav
+./dataset/44k/yuuka/380298.wav
+./dataset/44k/yuuka/1_56.wav
+./dataset/44k/yuuka/1_5.wav
+./dataset/44k/yuuka/1_94.wav
+./dataset/44k/yuuka/7_86.wav
+./dataset/44k/yuuka/7_42.wav
+./dataset/44k/yuuka/1_64.wav
+./dataset/44k/yuuka/5_115.wav
+./dataset/44k/yuuka/audio_26.wav
+./dataset/44k/yuuka/4_63.wav
+./dataset/44k/yuuka/5_83.wav
+./dataset/44k/yuuka/5_64.wav
+./dataset/44k/yuuka/7_150.wav
+./dataset/44k/yuuka/7_179.wav
+./dataset/44k/yuuka/6_35.wav
+./dataset/44k/yuuka/7_9.wav
+./dataset/44k/yuuka/3_127.wav
+./dataset/44k/yuuka/1_21.wav
+./dataset/44k/yuuka/5_31.wav
+./dataset/44k/yuuka/5_78.wav
+./dataset/44k/yuuka/9_6.wav
+./dataset/44k/yuuka/4_71.wav
+./dataset/44k/yuuka/2_9.wav
+./dataset/44k/yuuka/617112.wav
+./dataset/44k/yuuka/8_28.wav
+./dataset/44k/yuuka/5_11.wav
+./dataset/44k/yuuka/832708.wav
+./dataset/44k/yuuka/8_0.wav
+./dataset/44k/yuuka/540598.wav
+./dataset/44k/yuuka/7_38.wav
+./dataset/44k/yuuka/1_71.wav
+./dataset/44k/yuuka/6_37.wav
+./dataset/44k/yuuka/394815.wav
+./dataset/44k/yuuka/9_24.wav
+./dataset/44k/yuuka/5_49.wav
+./dataset/44k/yuuka/3_103.wav
+./dataset/44k/yuuka/1_90.wav
+./dataset/44k/yuuka/9_13.wav
+./dataset/44k/yuuka/66052.wav
+./dataset/44k/yuuka/1_96.wav
+./dataset/44k/yuuka/3_141.wav
+./dataset/44k/yuuka/1_40.wav
+./dataset/44k/yuuka/2_17.wav
+./dataset/44k/yuuka/7_88.wav
+./dataset/44k/yuuka/386393.wav
+./dataset/44k/yuuka/2_10.wav
+./dataset/44k/yuuka/4_79.wav
+./dataset/44k/yuuka/9_33.wav
+./dataset/44k/yuuka/7_91.wav
+./dataset/44k/yuuka/audio_33.wav
+./dataset/44k/yuuka/7_0.wav
+./dataset/44k/yuuka/3_138.wav
+./dataset/44k/yuuka/872587.wav
+./dataset/44k/yuuka/7_122.wav
+./dataset/44k/yuuka/40927.wav
+./dataset/44k/yuuka/5_33.wav
+./dataset/44k/yuuka/4_11.wav
+./dataset/44k/yuuka/4_40.wav
+./dataset/44k/yuuka/3_28.wav
+./dataset/44k/yuuka/audio_37.wav
+./dataset/44k/yuuka/228114.wav
+./dataset/44k/yuuka/629546.wav
+./dataset/44k/yuuka/1_100.wav
+./dataset/44k/yuuka/4_62.wav
+./dataset/44k/yuuka/1_6.wav
+./dataset/44k/yuuka/650930.wav
+./dataset/44k/yuuka/7_121.wav
+./dataset/44k/yuuka/1_95.wav
+./dataset/44k/yuuka/audio_20.wav
+./dataset/44k/yuuka/752079.wav
+./dataset/44k/yuuka/7_33.wav
+./dataset/44k/yuuka/1_99.wav
+./dataset/44k/yuuka/5_1.wav
+./dataset/44k/yuuka/571895.wav
+./dataset/44k/yuuka/3_8.wav
+./dataset/44k/yuuka/5_99.wav
+./dataset/44k/yuuka/5_14.wav
+./dataset/44k/yuuka/3_68.wav
+./dataset/44k/yuuka/7_161.wav
+./dataset/44k/yuuka/2_16.wav
+./dataset/44k/yuuka/audio_2.wav
+./dataset/44k/yuuka/2_4.wav
+./dataset/44k/yuuka/1_3.wav
+./dataset/44k/yuuka/260935.wav
+./dataset/44k/yuuka/3_152.wav
+./dataset/44k/yuuka/4_35.wav
+./dataset/44k/yuuka/617923.wav
+./dataset/44k/yuuka/5_98.wav
+./dataset/44k/yuuka/4_49.wav
+./dataset/44k/yuuka/audio_43.wav
+./dataset/44k/yuuka/73560.wav
+./dataset/44k/yuuka/3_144.wav
+./dataset/44k/yuuka/5_82.wav
+./dataset/44k/yuuka/4_7.wav
+./dataset/44k/yuuka/651496.wav
+./dataset/44k/yuuka/3_49.wav
+./dataset/44k/yuuka/audio_6.wav
+./dataset/44k/yuuka/9_71.wav
+./dataset/44k/yuuka/1_59.wav
+./dataset/44k/yuuka/3_25.wav
+./dataset/44k/yuuka/7_92.wav
+./dataset/44k/yuuka/7_46.wav
+./dataset/44k/yuuka/4_73.wav
+./dataset/44k/yuuka/7_136.wav
+./dataset/44k/yuuka/9_79.wav
+./dataset/44k/yuuka/3_175.wav
+./dataset/44k/yuuka/9_41.wav
+./dataset/44k/yuuka/9_31.wav
+./dataset/44k/yuuka/761979.wav
+./dataset/44k/yuuka/3_83.wav
+./dataset/44k/yuuka/3_177.wav
+./dataset/44k/yuuka/7_163.wav
+./dataset/44k/yuuka/3_43.wav
+./dataset/44k/yuuka/3_37.wav
+./dataset/44k/yuuka/8_2.wav
+./dataset/44k/yuuka/8_10.wav
+./dataset/44k/yuuka/820939.wav
+./dataset/44k/yuuka/778182.wav
+./dataset/44k/yuuka/6_31.wav
+./dataset/44k/yuuka/3_164.wav
+./dataset/44k/yuuka/5_32.wav
+./dataset/44k/yuuka/215854.wav
+./dataset/44k/yuuka/6_5.wav
+./dataset/44k/yuuka/3_109.wav
+./dataset/44k/yuuka/6_42.wav
+./dataset/44k/yuuka/188142.wav
+./dataset/44k/yuuka/6_40.wav
+./dataset/44k/yuuka/4_30.wav
+./dataset/44k/yuuka/3_161.wav
+./dataset/44k/yuuka/7_62.wav
+./dataset/44k/yuuka/2_22.wav
+./dataset/44k/yuuka/4_69.wav
+./dataset/44k/yuuka/4_6.wav
+./dataset/44k/yuuka/577791.wav
+./dataset/44k/yuuka/723735.wav
+./dataset/44k/yuuka/7_90.wav
+./dataset/44k/yuuka/5_96.wav
+./dataset/44k/yuuka/1_82.wav
+./dataset/44k/yuuka/225368.wav
+./dataset/44k/yuuka/audio_31.wav
+./dataset/44k/yuuka/3_130.wav
+./dataset/44k/yuuka/audio_39.wav
+./dataset/44k/yuuka/7_153.wav
+./dataset/44k/yuuka/7_183.wav
+./dataset/44k/yuuka/870812.wav
+./dataset/44k/yuuka/5_114.wav
+./dataset/44k/yuuka/719896.wav
+./dataset/44k/yuuka/5_65.wav
+./dataset/44k/yuuka/269297.wav
+./dataset/44k/yuuka/3_124.wav
+./dataset/44k/yuuka/9_57.wav
+./dataset/44k/yuuka/7_18.wav
+./dataset/44k/yuuka/619226.wav
+./dataset/44k/yuuka/207781.wav
+./dataset/44k/yuuka/5_28.wav
+./dataset/44k/yuuka/5_71.wav
+./dataset/44k/yuuka/143646.wav
+./dataset/44k/yuuka/383188.wav
+./dataset/44k/yuuka/7_131.wav
+./dataset/44k/yuuka/3_92.wav
+./dataset/44k/yuuka/892318.wav
+./dataset/44k/yuuka/1_57.wav
+./dataset/44k/yuuka/audio_21.wav
+./dataset/44k/yuuka/4_57.wav
+./dataset/44k/yuuka/3_96.wav
+./dataset/44k/yuuka/7_12.wav
+./dataset/44k/yuuka/6_29.wav
+./dataset/44k/yuuka/3_3.wav
+./dataset/44k/yuuka/7_97.wav
+./dataset/44k/yuuka/1_54.wav
+./dataset/44k/yuuka/9_78.wav
+./dataset/44k/yuuka/76879.wav
+./dataset/44k/yuuka/3_123.wav
+./dataset/44k/yuuka/1_91.wav
+./dataset/44k/yuuka/7_138.wav
+./dataset/44k/yuuka/7_77.wav
+./dataset/44k/yuuka/5_55.wav
+./dataset/44k/yuuka/2_29.wav
+./dataset/44k/yuuka/2_19.wav
+./dataset/44k/yuuka/4_45.wav
+./dataset/44k/yuuka/7_84.wav
+./dataset/44k/yuuka/3_143.wav
+./dataset/44k/yuuka/1_8.wav
+./dataset/44k/yuuka/3_64.wav
+./dataset/44k/yuuka/1_30.wav
+./dataset/44k/yuuka/3_10.wav
+./dataset/44k/yuuka/854231.wav
+./dataset/44k/yuuka/6_26.wav
+./dataset/44k/yuuka/3_23.wav
+./dataset/44k/yuuka/5_44.wav
+./dataset/44k/yuuka/7_45.wav
+./dataset/44k/yuuka/5_10.wav
+./dataset/44k/yuuka/7_180.wav
+./dataset/44k/yuuka/9_12.wav
+./dataset/44k/yuuka/audio_8.wav
+./dataset/44k/yuuka/3_35.wav
+./dataset/44k/yuuka/565217.wav
+./dataset/44k/yuuka/7_63.wav
+./dataset/44k/yuuka/5_75.wav
+./dataset/44k/yuuka/1_2.wav
+./dataset/44k/yuuka/9_8.wav
+./dataset/44k/yuuka/5_94.wav
+./dataset/44k/yuuka/1_75.wav
+./dataset/44k/yuuka/8_8.wav
+./dataset/44k/yuuka/7_141.wav
+./dataset/44k/yuuka/5_17.wav
+./dataset/44k/yuuka/853514.wav
+./dataset/44k/yuuka/4_78.wav
+./dataset/44k/yuuka/5_116.wav
+./dataset/44k/yuuka/448774.wav
+./dataset/44k/yuuka/8_13.wav
+./dataset/44k/yuuka/4_13.wav
+./dataset/44k/yuuka/4_8.wav
+./dataset/44k/yuuka/5_60.wav
+./dataset/44k/yuuka/321094.wav
+./dataset/44k/yuuka/859348.wav
+./dataset/44k/yuuka/9_19.wav
+./dataset/44k/yuuka/1_27.wav
+./dataset/44k/yuuka/4_58.wav
+./dataset/44k/yuuka/1_53.wav
+./dataset/44k/yuuka/522647.wav
+./dataset/44k/yuuka/73298.wav
+./dataset/44k/yuuka/9_2.wav
+./dataset/44k/yuuka/3_139.wav
+./dataset/44k/yuuka/9_55.wav
+./dataset/44k/yuuka/1_115.wav
+./dataset/44k/yuuka/796714.wav
+./dataset/44k/yuuka/5_43.wav
+./dataset/44k/yuuka/3_57.wav
+./dataset/44k/yuuka/4_70.wav
+./dataset/44k/yuuka/4_81.wav
+./dataset/44k/yuuka/4_31.wav
+./dataset/44k/yuuka/3_88.wav
+./dataset/44k/yuuka/5_57.wav
+./dataset/44k/yuuka/6_19.wav
+./dataset/44k/yuuka/6_10.wav
+./dataset/44k/yuuka/9_20.wav
+./dataset/44k/yuuka/5_37.wav
+./dataset/44k/yuuka/6_43.wav
+./dataset/44k/yuuka/9_23.wav
+./dataset/44k/yuuka/7_60.wav
+./dataset/44k/yuuka/3_159.wav
+./dataset/44k/yuuka/182426.wav
+./dataset/44k/yuuka/1_50.wav
+./dataset/44k/yuuka/5_81.wav
+./dataset/44k/yuuka/4_64.wav
+./dataset/44k/yuuka/5_84.wav
+./dataset/44k/yuuka/5_29.wav
+./dataset/44k/yuuka/2_8.wav
+./dataset/44k/yuuka/4_32.wav
+./dataset/44k/yuuka/7_47.wav
+./dataset/44k/yuuka/7_109.wav
+./dataset/44k/yuuka/130059.wav
+./dataset/44k/yuuka/1_42.wav
+./dataset/44k/yuuka/5_47.wav
+./dataset/44k/yuuka/9_40.wav
+./dataset/44k/yuuka/4_75.wav
+./dataset/44k/yuuka/1_63.wav
+./dataset/44k/yuuka/9_9.wav
+./dataset/44k/yuuka/208859.wav
+./dataset/44k/yuuka/5_56.wav
+./dataset/44k/yuuka/9_36.wav
+./dataset/44k/yuuka/187864.wav
+./dataset/44k/yuuka/7_21.wav
+./dataset/44k/yuuka/7_144.wav
+./dataset/44k/yuuka/7_22.wav
+./dataset/44k/yuuka/7_34.wav
+./dataset/44k/yuuka/1_65.wav
+./dataset/44k/yuuka/6_28.wav
+./dataset/44k/yuuka/7_19.wav

filelists/val.txt

@@ -0,0 +1,2 @@
+./dataset/44k/yuuka/3_106.wav
+./dataset/44k/yuuka/3_7.wav

flask_api.py

@@ -0,0 +1,57 @@
+import io
+import logging
+
+import librosa
+import soundfile
+import paddle
+import paddle.audio as paddleaudio
+from flask import Flask, request, send_file
+from flask_cors import CORS
+
+from inference.infer_tool import Svc, RealTimeVC
+
+app = Flask(__name__)
+
+CORS(app)
+
+logging.getLogger('numba').setLevel(logging.WARNING)
+
+
+@app.route("/voiceChangeModel", methods=["POST"])
+def voice_change_model():
+    request_form = request.form
+    wave_file = request.files.get("sample", None)
+    # 变调信息
+    f_pitch_change = float(request_form.get("fPitchChange", 0))
+    # DAW所需的采样率
+    daw_sample = int(float(request_form.get("sampleRate", 0)))
+    speaker_id = int(float(request_form.get("sSpeakId", 0)))
+    # http获得wav文件并转换
+    input_wav_path = io.BytesIO(wave_file.read())
+
+    # 模型推理
+    if raw_infer:
+        out_audio, out_sr = svc_model.infer(speaker_id, f_pitch_change, input_wav_path)
+        tar_audio = librosa.resample(out_audio.numpy(), svc_model.target_sample, daw_sample) 
+    else:
+        out_audio = svc.process(svc_model, speaker_id, f_pitch_change, input_wav_path)
+        tar_audio = librosa.resample(out_audio, svc_model.target_sample, daw_sample)
+    # 返回音频
+    out_wav_path = io.BytesIO()
+    soundfile.write(out_wav_path, tar_audio, daw_sample, format="wav")
+    out_wav_path.seek(0)
+    return send_file(out_wav_path, download_name="temp.wav", as_attachment=True)
+
+
+if __name__ == '__main__':
+    # 启用则为直接切片合成，False为交叉淡化方式
+    # vst插件调整0.3-0.5s切片时间可以降低延迟，直接切片方法会有连接处爆音、交叉淡化会有轻微重叠声音
+    # 自行选择能接受的方法，或将vst最大切片时间调整为1s，此处设为Ture，延迟大音质稳定一些
+    raw_infer = True
+    # 每个模型和config是唯一对应的
+    model_name = "logs/44k/G_1005.pdparams"
+    config_name = "configs/config.json"
+    svc_model = Svc(model_name, config_name)
+    svc = RealTimeVC()
+    # 此处与vst插件对应，不建议更改
+    app.run(port=6842, host="0.0.0.0", debug=False, threaded=False)

hubert/hubert4.0.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cc348818d386f8cff3332ab4bcd77e5870c373c492c896664092fd7230122a32
+size 293347531

hubert/hubert_model.py

@@ -0,0 +1,226 @@
+import copy
+import random
+from typing import Optional, Tuple
+
+import paddle
+import paddle.nn as nn
+import paddle.nn.functional as t_func
+#from torch.nn.modules.utils import consume_prefix_in_state_dict_if_present
+
+
+class Hubert(paddle.nn.Layer):
+    def __init__(self, num_label_embeddings: int = 100, mask: bool = True):
+        super().__init__()
+        self._mask = mask
+        self.feature_extractor = FeatureExtractor()
+        self.feature_projection = FeatureProjection()
+        self.positional_embedding = PositionalConvEmbedding()
+        self.norm = nn.LayerNorm(768)
+        self.dropout = nn.Dropout(0.1)
+        self.encoder = TransformerEncoder(
+            nn.TransformerEncoderLayer(
+                768, 12, 3072, activation="gelu"
+            ),
+            12,
+        )
+        self.proj = nn.Linear(768, 256)
+
+        self.masked_spec_embed = paddle.create_parameter([768],dtype = 'float32')
+        self.label_embedding = nn.Embedding(num_label_embeddings, 256)
+
+    def mask(self, x: paddle.Tensor) -> Tuple[paddle.Tensor, paddle.Tensor]:
+        mask = None
+        if self.training and self._mask:
+            mask = _compute_mask((x.size(0), x.size(1)), 0.8, 10, None, 2)
+            x[mask] = self.masked_spec_embed
+        return x, mask
+
+    def encode(
+            self, x: paddle.Tensor, layer: Optional[int] = None
+    ) -> Tuple[paddle.Tensor, paddle.Tensor]:
+        x = self.feature_extractor(x)
+        x = self.feature_projection(x.transpose([0, 2, 1]))
+        x, mask = self.mask(x)
+        x = x + self.positional_embedding(x)
+        x = self.dropout(self.norm(x))
+        x = self.encoder(x, output_layer=layer)
+        return x, mask
+
+    def logits(self, x: paddle.Tensor) -> paddle.Tensor:
+        logits = t_func.cosine_similarity(
+            x.unsqueeze(2),
+            self.label_embedding.weight.unsqueeze(0).unsqueeze(0),
+            axis=-1,
+        )
+        return logits / 0.1
+
+    def forward(self, x: paddle.Tensor) -> Tuple[paddle.Tensor, paddle.Tensor]:
+        x, mask = self.encode(x)
+        x = self.proj(x)
+        logits = self.logits(x)
+        return logits, mask
+
+
+class HubertSoft(Hubert):
+    def __init__(self):
+        super().__init__()
+
+    def units(self, wav: paddle.Tensor) -> paddle.Tensor:
+        wav = t_func.pad(wav, ((400 - 320) // 2, (400 - 320) // 2),data_format='NCL')
+        x, _ = self.encode(wav)
+        return self.proj(x)
+
+
+class FeatureExtractor(paddle.nn.Layer):
+    def __init__(self):
+        super().__init__()
+        self.conv0 = nn.Conv1D(1, 512, 10, 5, bias_attr=False)
+        self.norm0 = nn.GroupNorm(512, 512)
+        self.conv1 = nn.Conv1D(512, 512, 3, 2, bias_attr=False)
+        self.conv2 = nn.Conv1D(512, 512, 3, 2, bias_attr=False)
+        self.conv3 = nn.Conv1D(512, 512, 3, 2, bias_attr=False)
+        self.conv4 = nn.Conv1D(512, 512, 3, 2, bias_attr=False)
+        self.conv5 = nn.Conv1D(512, 512, 2, 2, bias_attr=False)
+        self.conv6 = nn.Conv1D(512, 512, 2, 2, bias_attr=False)
+
+    def forward(self, x: paddle.Tensor) -> paddle.Tensor:
+        x = t_func.gelu(self.norm0(self.conv0(x)))
+        x = t_func.gelu(self.conv1(x))
+        x = t_func.gelu(self.conv2(x))
+        x = t_func.gelu(self.conv3(x))
+        x = t_func.gelu(self.conv4(x))
+        x = t_func.gelu(self.conv5(x))
+        x = t_func.gelu(self.conv6(x))
+        return x
+
+
+class FeatureProjection(paddle.nn.Layer):
+    def __init__(self):
+        super().__init__()
+        self.norm = nn.LayerNorm(512)
+        self.projection = nn.Linear(512, 768)
+        self.dropout = nn.Dropout(0.1)
+
+    def forward(self, x: paddle.Tensor) -> paddle.Tensor:
+        x = self.norm(x)
+        x = self.projection(x)
+        x = self.dropout(x)
+        return x
+
+
+class PositionalConvEmbedding(paddle.nn.Layer):
+    def __init__(self):
+        super().__init__()
+        self.conv = nn.Conv1D(
+            768,
+            768,
+            kernel_size=128,
+            padding=128 // 2,
+            groups=16,
+        )
+        self.conv = nn.utils.weight_norm(self.conv, name="weight", dim=2)
+
+    def forward(self, x: paddle.Tensor) -> paddle.Tensor:
+        x = self.conv(x.transpose([0, 2, 1]))
+        x = t_func.gelu(x[:, :, :-1])
+        return x.transpose([0, 2, 1])
+
+
+class TransformerEncoder(paddle.nn.Layer):
+    def __init__(
+            self, encoder_layer: nn.TransformerEncoderLayer, num_layers: int
+    ) -> None:
+        super(TransformerEncoder, self).__init__()
+        self.layers = nn.LayerList(
+            [copy.deepcopy(encoder_layer) for _ in range(num_layers)]
+        )
+        self.num_layers = num_layers
+
+    def forward(
+            self,
+            src: paddle.Tensor,
+            mask: paddle.Tensor = None,
+            src_key_padding_mask: paddle.Tensor = None,
+            output_layer: Optional[int] = None,
+    ) -> paddle.Tensor:
+        output = src
+        for layer in self.layers[:output_layer]:
+            output = layer(
+                output, src_mask=mask, src_key_padding_mask=src_key_padding_mask
+            )
+        return output
+
+
+def _compute_mask(
+        shape: Tuple[int, int],
+        mask_prob: float,
+        mask_length: int,
+        device: None,
+        min_masks: int = 0,
+) -> paddle.Tensor:
+    batch_size, sequence_length = shape
+
+    if mask_length < 1:
+        raise ValueError("`mask_length` has to be bigger than 0.")
+
+    if mask_length > sequence_length:
+        raise ValueError(
+            f"`mask_length` has to be smaller than `sequence_length`, but got `mask_length`: {mask_length} and `sequence_length`: {sequence_length}`"
+        )
+
+    # compute number of masked spans in batch
+    num_masked_spans = int(mask_prob * sequence_length / mask_length + random.random())
+    num_masked_spans = max(num_masked_spans, min_masks)
+
+    # make sure num masked indices <= sequence_length
+    if num_masked_spans * mask_length > sequence_length:
+        num_masked_spans = sequence_length // mask_length
+
+    # SpecAugment mask to fill
+    mask = paddle.zeros((batch_size, sequence_length), dtype='bool')
+
+    # uniform distribution to sample from, make sure that offset samples are < sequence_length
+    uniform_dist = paddle.ones(
+        (batch_size, sequence_length - (mask_length - 1))
+    )
+    
+    # get random indices to mask
+    mask_indices = paddle.multinomial(uniform_dist, num_masked_spans)
+
+    # expand masked indices to masked spans
+    mask_indices = (
+        mask_indices.unsqueeze(dim=-1)
+        .expand((batch_size, num_masked_spans, mask_length))
+        .reshape(batch_size, num_masked_spans * mask_length)
+    )
+    offsets = (
+        paddle.arange(mask_length)[None, None, :]
+        .expand((batch_size, num_masked_spans, mask_length))
+        .reshape(batch_size, num_masked_spans * mask_length)
+    )
+    mask_idxs = mask_indices + offsets
+
+    # scatter indices to mask
+    mask = mask.scatter(1, mask_idxs, True)
+
+    return mask
+
+
+def hubert_soft(
+        path: str,
+) -> HubertSoft:
+    r"""HuBERT-Soft from `"A Comparison of Discrete and Soft Speech Units for Improved Voice Conversion"`.
+    Args:
+        path (str): path of a pretrained model
+    """
+    hubert = HubertSoft()
+    checkpoint = paddle.load(path)
+    #consume_prefix_in_state_dict_if_present(checkpoint, "module.")
+    hubert.set_state_dict(checkpoint)
+    hubert.eval()
+    return hubert
+
+if __name__ == '__main__':
+    hubert = HubertSoft()
+    d = paddle.load(r'E:\trans\hubert\final.pdparams')
+    hubert.set_state_dict(d)

hubert/hubert_model_onnx.py

@@ -0,0 +1,217 @@
+import copy
+import random
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as t_func
+from torch.nn.modules.utils import consume_prefix_in_state_dict_if_present
+
+
+class Hubert(nn.Layer):
+    def __init__(self, num_label_embeddings: int = 100, mask: bool = True):
+        super().__init__()
+        self._mask = mask
+        self.feature_extractor = FeatureExtractor()
+        self.feature_projection = FeatureProjection()
+        self.positional_embedding = PositionalConvEmbedding()
+        self.norm = nn.LayerNorm(768)
+        self.dropout = nn.Dropout(0.1)
+        self.encoder = TransformerEncoder(
+            nn.TransformerEncoderLayer(
+                768, 12, 3072, activation="gelu", batch_first=True
+            ),
+            12,
+        )
+        self.proj = nn.Linear(768, 256)
+
+        self.masked_spec_embed = nn.Parameter(torch.FloatTensor(768).uniform_())
+        self.label_embedding = nn.Embedding(num_label_embeddings, 256)
+
+    def mask(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+        mask = None
+        if self.training and self._mask:
+            mask = _compute_mask((x.size(0), x.size(1)), 0.8, 10, x.device, 2)
+            x[mask] = self.masked_spec_embed.to(x.dtype)
+        return x, mask
+
+    def encode(
+            self, x: torch.Tensor, layer: Optional[int] = None
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        x = self.feature_extractor(x)
+        x = self.feature_projection(x.transpose(1, 2))
+        x, mask = self.mask(x)
+        x = x + self.positional_embedding(x)
+        x = self.dropout(self.norm(x))
+        x = self.encoder(x, output_layer=layer)
+        return x, mask
+
+    def logits(self, x: torch.Tensor) -> torch.Tensor:
+        logits = torch.cosine_similarity(
+            x.unsqueeze(2),
+            self.label_embedding.weight.unsqueeze(0).unsqueeze(0),
+            dim=-1,
+        )
+        return logits / 0.1
+
+
+class HubertSoft(Hubert):
+    def __init__(self):
+        super().__init__()
+
+    def units(self, wav: torch.Tensor) -> torch.Tensor:
+        wav = t_func.pad(wav, ((400 - 320) // 2, (400 - 320) // 2))
+        x, _ = self.encode(wav)
+        return self.proj(x)
+
+    def forward(self, x):
+        return self.units(x)
+
+class FeatureExtractor(nn.Layer):
+    def __init__(self):
+        super().__init__()
+        self.conv0 = nn.Conv1d(1, 512, 10, 5, bias=False)
+        self.norm0 = nn.GroupNorm(512, 512)
+        self.conv1 = nn.Conv1d(512, 512, 3, 2, bias=False)
+        self.conv2 = nn.Conv1d(512, 512, 3, 2, bias=False)
+        self.conv3 = nn.Conv1d(512, 512, 3, 2, bias=False)
+        self.conv4 = nn.Conv1d(512, 512, 3, 2, bias=False)
+        self.conv5 = nn.Conv1d(512, 512, 2, 2, bias=False)
+        self.conv6 = nn.Conv1d(512, 512, 2, 2, bias=False)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = t_func.gelu(self.norm0(self.conv0(x)))
+        x = t_func.gelu(self.conv1(x))
+        x = t_func.gelu(self.conv2(x))
+        x = t_func.gelu(self.conv3(x))
+        x = t_func.gelu(self.conv4(x))
+        x = t_func.gelu(self.conv5(x))
+        x = t_func.gelu(self.conv6(x))
+        return x
+
+
+class FeatureProjection(nn.Layer):
+    def __init__(self):
+        super().__init__()
+        self.norm = nn.LayerNorm(512)
+        self.projection = nn.Linear(512, 768)
+        self.dropout = nn.Dropout(0.1)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.norm(x)
+        x = self.projection(x)
+        x = self.dropout(x)
+        return x
+
+
+class PositionalConvEmbedding(nn.Layer):
+    def __init__(self):
+        super().__init__()
+        self.conv = nn.Conv1d(
+            768,
+            768,
+            kernel_size=128,
+            padding=128 // 2,
+            groups=16,
+        )
+        self.conv = nn.utils.weight_norm(self.conv, name="weight", dim=2)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.conv(x.transpose(1, 2))
+        x = t_func.gelu(x[:, :, :-1])
+        return x.transpose(1, 2)
+
+
+class TransformerEncoder(nn.Layer):
+    def __init__(
+            self, encoder_layer: nn.TransformerEncoderLayer, num_layers: int
+    ) -> None:
+        super(TransformerEncoder, self).__init__()
+        self.layers = nn.LayerList(
+            [copy.deepcopy(encoder_layer) for _ in range(num_layers)]
+        )
+        self.num_layers = num_layers
+
+    def forward(
+            self,
+            src: torch.Tensor,
+            mask: torch.Tensor = None,
+            src_key_padding_mask: torch.Tensor = None,
+            output_layer: Optional[int] = None,
+    ) -> torch.Tensor:
+        output = src
+        for layer in self.layers[:output_layer]:
+            output = layer(
+                output, src_mask=mask, src_key_padding_mask=src_key_padding_mask
+            )
+        return output
+
+
+def _compute_mask(
+        shape: Tuple[int, int],
+        mask_prob: float,
+        mask_length: int,
+        device: torch.device,
+        min_masks: int = 0,
+) -> torch.Tensor:
+    batch_size, sequence_length = shape
+
+    if mask_length < 1:
+        raise ValueError("`mask_length` has to be bigger than 0.")
+
+    if mask_length > sequence_length:
+        raise ValueError(
+            f"`mask_length` has to be smaller than `sequence_length`, but got `mask_length`: {mask_length} and `sequence_length`: {sequence_length}`"
+        )
+
+    # compute number of masked spans in batch
+    num_masked_spans = int(mask_prob * sequence_length / mask_length + random.random())
+    num_masked_spans = max(num_masked_spans, min_masks)
+
+    # make sure num masked indices <= sequence_length
+    if num_masked_spans * mask_length > sequence_length:
+        num_masked_spans = sequence_length // mask_length
+
+    # SpecAugment mask to fill
+    mask = torch.zeros((batch_size, sequence_length), device=device, dtype=torch.bool)
+
+    # uniform distribution to sample from, make sure that offset samples are < sequence_length
+    uniform_dist = torch.ones(
+        (batch_size, sequence_length - (mask_length - 1)), device=device
+    )
+
+    # get random indices to mask
+    mask_indices = torch.multinomial(uniform_dist, num_masked_spans)
+
+    # expand masked indices to masked spans
+    mask_indices = (
+        mask_indices.unsqueeze(dim=-1)
+        .expand((batch_size, num_masked_spans, mask_length))
+        .reshape(batch_size, num_masked_spans * mask_length)
+    )
+    offsets = (
+        torch.arange(mask_length, device=device)[None, None, :]
+        .expand((batch_size, num_masked_spans, mask_length))
+        .reshape(batch_size, num_masked_spans * mask_length)
+    )
+    mask_idxs = mask_indices + offsets
+
+    # scatter indices to mask
+    mask = mask.scatter(1, mask_idxs, True)
+
+    return mask
+
+
+def hubert_soft(
+        path: str,
+) -> HubertSoft:
+    r"""HuBERT-Soft from `"A Comparison of Discrete and Soft Speech Units for Improved Voice Conversion"`.
+    Args:
+        path (str): path of a pretrained model
+    """
+    hubert = HubertSoft()
+    checkpoint = torch.load(path)
+    consume_prefix_in_state_dict_if_present(checkpoint, "module.")
+    hubert.load_state_dict(checkpoint)
+    hubert.eval()
+    return hubert

inference/__init__.py

@@ -0,0 +1 @@
+'''梅花三弄再回首花了一个小时迁移的模块'''

inference/chunks_temp.json

@@ -0,0 +1 @@
+{}

inference/infer_tool.py

@@ -0,0 +1,255 @@
+import hashlib
+import io
+import json
+import logging
+import os
+import time
+from pathlib import Path
+from inference import slicer
+
+import librosa
+import numpy as np
+# import onnxruntime
+import parselmouth
+import soundfile
+import paddle
+import paddle.audio as paddleaudio
+import paddleaudio
+
+import cluster
+#from hubert import hubert_model
+import utils
+from models import SynthesizerTrn,SynthesizerTrn_test
+
+logging.getLogger('matplotlib').setLevel(logging.WARNING)
+paddle.audio.backends.set_backend('soundfile')
+
+def read_temp(file_name):
+    if not os.path.exists(file_name):
+        with open(file_name, "w") as f:
+            f.write(json.dumps({"info": "temp_dict"}))
+        return {}
+    else:
+        try:
+            with open(file_name, "r") as f:
+                data = f.read()
+            data_dict = json.loads(data)
+            if os.path.getsize(file_name) > 50 * 1024 * 1024:
+                f_name = file_name.replace("\\", "/").split("/")[-1]
+                print(f"clean {f_name}")
+                for wav_hash in list(data_dict.keys()):
+                    if int(time.time()) - int(data_dict[wav_hash]["time"]) > 14 * 24 * 3600:
+                        del data_dict[wav_hash]
+        except Exception as e:
+            print(e)
+            print(f"{file_name} error,auto rebuild file")
+            data_dict = {"info": "temp_dict"}
+        return data_dict
+
+
+def write_temp(file_name, data):
+    with open(file_name, "w") as f:
+        f.write(json.dumps(data))
+
+
+def timeit(func):
+    def run(*args, **kwargs):
+        t = time.time()
+        res = func(*args, **kwargs)
+        print('executing \'%s\' costed %.3fs' % (func.__name__, time.time() - t))
+        return res
+
+    return run
+
+
+def format_wav(audio_path):
+    if Path(audio_path).suffix == '.wav':
+        return
+    raw_audio, raw_sample_rate = librosa.load(audio_path, mono=True, sr=None)
+    soundfile.write(Path(audio_path).with_suffix(".wav"), raw_audio, raw_sample_rate)
+
+
+def get_end_file(dir_path, end):
+    file_lists = []
+    for root, dirs, files in os.walk(dir_path):
+        files = [f for f in files if f[0] != '.']
+        dirs[:] = [d for d in dirs if d[0] != '.']
+        for f_file in files:
+            if f_file.endswith(end):
+                file_lists.append(os.path.join(root, f_file).replace("\\", "/"))
+    return file_lists
+
+
+def get_md5(content):
+    return hashlib.new("md5", content).hexdigest()
+
+def fill_a_to_b(a, b):
+    if len(a) < len(b):
+        for _ in range(0, len(b) - len(a)):
+            a.append(a[0])
+
+def mkdir(paths: list):
+    for path in paths:
+        if not os.path.exists(path):
+            os.mkdir(path)
+
+def pad_array(arr, target_length):
+    current_length = arr.shape[0]
+    if current_length >= target_length:
+        return arr
+    else:
+        pad_width = target_length - current_length
+        pad_left = pad_width // 2
+        pad_right = pad_width - pad_left
+        padded_arr = np.pad(arr, (pad_left, pad_right), 'constant', constant_values=(0, 0))
+        return padded_arr
+
+
+class Svc(object):
+    def __init__(self, net_g_path, config_path,
+                 device=None,
+                 cluster_model_path="./logs/44k/kmeans_10000.pdparams",mode="train"):
+        self.net_g_path = net_g_path
+        if device is None:
+            self.dev = "gpu:0" if paddle.device.is_compiled_with_cuda() else "cpu"
+        else:
+            self.dev = device
+        self.net_g_ms = None
+        self.hps_ms = utils.get_hparams_from_file(config_path)
+        self.target_sample = self.hps_ms.data.sampling_rate
+        self.hop_size = self.hps_ms.data.hop_length
+        self.spk2id = self.hps_ms.spk
+        # 加载hubert
+        self.hubert_model = utils.get_hubert_model()
+        self.load_model(mode)
+        if os.path.exists(cluster_model_path):
+            self.cluster_model = cluster.get_cluster_model(cluster_model_path)
+
+    def load_model(self,mode):
+        # 获取模型配置
+        if mode == "train":
+            self.net_g_ms = SynthesizerTrn(
+                self.hps_ms.data.filter_length // 2 + 1,
+                self.hps_ms.train.segment_size // self.hps_ms.data.hop_length,
+                **self.hps_ms.model)
+        elif mode == "test":
+            self.net_g_ms = SynthesizerTrn_test(
+                self.hps_ms.data.filter_length // 2 + 1,
+                self.hps_ms.train.segment_size // self.hps_ms.data.hop_length,
+                **self.hps_ms.model)
+        _ = utils.load_checkpoint(self.net_g_path, self.net_g_ms, None)
+        if "half" in self.net_g_path and paddle.device.is_compiled_with_cuda():
+            self.net_g_ms.half().eval()
+            self.net_g_ms.half().to(self.dev)
+        else:
+            self.net_g_ms.eval()
+            self.net_g_ms.to(self.dev)
+
+
+
+    def get_unit_f0(self, in_path, tran, cluster_infer_ratio, speaker):
+
+        wav, sr = librosa.load(in_path, sr=self.target_sample)
+
+        f0 = utils.compute_f0_parselmouth(wav, sampling_rate=self.target_sample, hop_length=self.hop_size)
+        f0, uv = utils.interpolate_f0(f0)
+        f0 = paddle.to_tensor(f0,dtype = ('float32'))
+        uv = paddle.to_tensor(uv,dtype = ('float32'))
+        f0 = f0 * 2 ** (tran / 12)
+        f0 = f0.unsqueeze(0)
+        uv = uv.unsqueeze(0)
+
+        wav16k = librosa.resample(wav, orig_sr=self.target_sample, target_sr=16000)
+        wav16k = paddle.to_tensor(wav16k)
+        c = utils.get_hubert_content(self.hubert_model, wav_16k_tensor=wav16k)
+        c = utils.repeat_expand_2d(c.squeeze(0), f0.shape[1])
+
+        if cluster_infer_ratio !=0:
+            cluster_c = cluster.get_cluster_center_result(self.cluster_model, c.cpu().numpy().T, speaker).T
+            cluster_c = paddle.to_tensor(cluster_c,dtype = 'float32')
+            c = cluster_infer_ratio * cluster_c + (1 - cluster_infer_ratio) * c
+
+        c = c.unsqueeze(0)
+        return c, f0, uv
+
+    def infer(self, speaker, tran, raw_path,
+              cluster_infer_ratio=0,
+              auto_predict_f0=False,
+              noice_scale=0.4):
+        speaker_id = 0
+        sid = paddle.to_tensor([int(speaker_id)],dtype = 'int64').unsqueeze(0)
+        c, f0, uv = self.get_unit_f0(raw_path, tran, cluster_infer_ratio, speaker)
+        if "half" in self.net_g_path and paddle.device.is_compiled_with_cuda():
+            c = c.half()
+        with paddle.no_grad():
+            start = time.time()
+            audio = self.net_g_ms.infer(c, f0=f0, g=sid, uv=uv, predict_f0=auto_predict_f0, noice_scale=noice_scale)[0,0].detach().astype('float32')
+            use_time = time.time() - start
+            print("vits耗时：{}".format(use_time))
+        return audio, audio.shape[-1]
+
+    def slice_inference(self,raw_audio_path, spk, tran, slice_db,cluster_infer_ratio, auto_predict_f0,noice_scale, pad_seconds=0.5,empty_cache=False):
+        wav_path = raw_audio_path
+        chunks = slicer.cut(wav_path, db_thresh=slice_db)
+        audio_data, audio_sr = slicer.chunks2audio(wav_path, chunks)
+
+        audio = []
+        for (slice_tag, data) in audio_data:
+            print(f'#=====分段开始，耗时{round(len(data) / audio_sr, 3)}秒======')
+            # padd
+            pad_len = int(audio_sr * pad_seconds)
+            data = np.concatenate([np.zeros([pad_len]), data, np.zeros([pad_len])])
+            length = int(np.ceil(len(data) / audio_sr * self.target_sample))
+            raw_path = io.BytesIO()
+            soundfile.write(raw_path, data, audio_sr, format="wav")
+            raw_path.seek(0)
+            if slice_tag:
+                print('跳过空段')
+                _audio = np.zeros(length)
+            else:
+                out_audio, out_sr = self.infer(spk, tran, raw_path,
+                                                    cluster_infer_ratio=cluster_infer_ratio,
+                                                    auto_predict_f0=auto_predict_f0,
+                                                    noice_scale=noice_scale
+                                                    )
+                _audio = out_audio.cpu().numpy()
+
+            pad_len = int(self.target_sample * pad_seconds)
+            _audio = _audio[pad_len:-pad_len]
+            audio.extend(list(_audio))
+        if empty_cache == True:
+            paddle.device.cuda.empty_cache()
+        return np.array(audio)
+
+
+class RealTimeVC:
+    def __init__(self):
+        self.last_chunk = None
+        self.last_o = None
+        self.chunk_len = 16000  # 区块长度
+        self.pre_len = 3840  # 交叉淡化长度，640的倍数
+
+    """输入输出都是1维numpy 音频波形数组"""
+
+    def process(self, svc_model, speaker_id, f_pitch_change, input_wav_path):
+        import maad
+        audio, sr = paddleaudio.load(input_wav_path)
+        audio = audio.cpu().numpy()[0]
+        temp_wav = io.BytesIO()
+        if self.last_chunk is None:
+            input_wav_path.seek(0)
+            audio, sr = svc_model.infer(speaker_id, f_pitch_change, input_wav_path)
+            audio = audio.cpu().numpy()
+            self.last_chunk = audio[-self.pre_len:]
+            self.last_o = audio
+            return audio[-self.chunk_len:]
+        else:
+            audio = np.concatenate([self.last_chunk, audio])
+            soundfile.write(temp_wav, audio, sr, format="wav")
+            temp_wav.seek(0)
+            audio, sr = svc_model.infer(speaker_id, f_pitch_change, temp_wav)
+            audio = audio.cpu().numpy()
+            ret = maad.util.crossfade(self.last_o, audio, self.pre_len)
+            self.last_chunk = audio[-self.pre_len:]
+            self.last_o = audio
+            return ret[self.chunk_len:2 * self.chunk_len]

inference/infer_tool_grad.py

@@ -0,0 +1,161 @@
+import hashlib
+import json
+import logging
+import os
+import time
+from pathlib import Path
+import io
+import librosa
+import maad
+import numpy as np
+from inference import slicer
+import parselmouth
+import soundfile
+import paddle
+import paddle.audio as paddleaudio
+
+from hubert import hubert_model
+import utils
+from models import SynthesizerTrn
+logging.getLogger('numba').setLevel(logging.WARNING)
+logging.getLogger('matplotlib').setLevel(logging.WARNING)
+
+def resize2d_f0(x, target_len):
+    source = np.array(x)
+    source[source < 0.001] = np.nan
+    target = np.interp(np.arange(0, len(source) * target_len, len(source)) / target_len, np.arange(0, len(source)),
+                       source)
+    res = np.nan_to_num(target)
+    return res
+
+def get_f0(x, p_len,f0_up_key=0):
+
+    time_step = 160 / 16000 * 1000
+    f0_min = 50
+    f0_max = 1100
+    f0_mel_min = 1127 * np.log(1 + f0_min / 700)
+    f0_mel_max = 1127 * np.log(1 + f0_max / 700)
+
+    f0 = parselmouth.Sound(x, 16000).to_pitch_ac(
+        time_step=time_step / 1000, voicing_threshold=0.6,
+        pitch_floor=f0_min, pitch_ceiling=f0_max).selected_array['frequency']
+
+    pad_size=(p_len - len(f0) + 1) // 2
+    if(pad_size>0 or p_len - len(f0) - pad_size>0):
+        f0 = np.pad(f0,[[pad_size,p_len - len(f0) - pad_size]], mode='constant')
+
+    f0 *= pow(2, f0_up_key / 12)
+    f0_mel = 1127 * np.log(1 + f0 / 700)
+    f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - f0_mel_min) * 254 / (f0_mel_max - f0_mel_min) + 1
+    f0_mel[f0_mel <= 1] = 1
+    f0_mel[f0_mel > 255] = 255
+    f0_coarse = np.rint(f0_mel).astype(np.int)
+    return f0_coarse, f0
+
+def clean_pitch(input_pitch):
+    num_nan = np.sum(input_pitch == 1)
+    if num_nan / len(input_pitch) > 0.9:
+        input_pitch[input_pitch != 1] = 1
+    return input_pitch
+
+
+def plt_pitch(input_pitch):
+    input_pitch = input_pitch.astype(float)
+    input_pitch[input_pitch == 1] = np.nan
+    return input_pitch
+
+
+def f0_to_pitch(ff):
+    f0_pitch = 69 + 12 * np.log2(ff / 440)
+    return f0_pitch
+
+
+def fill_a_to_b(a, b):
+    if len(a) < len(b):
+        for _ in range(0, len(b) - len(a)):
+            a.append(a[0])
+
+
+def mkdir(paths: list):
+    for path in paths:
+        if not os.path.exists(path):
+            os.mkdir(path)
+
+
+class VitsSvc(object):
+    def __init__(self):
+        self.device = "gpu:0" if paddle.device.is_compiled_with_cuda() else "cpu"
+        self.SVCVITS = None
+        self.hps = None
+        self.speakers = None
+        self.hubert_soft = utils.get_hubert_model()
+
+    def set_device(self, device):
+        self.device = device
+        #self.hubert_soft.to(self.device)
+        if self.SVCVITS != None:
+            self.SVCVITS.to(self.device)
+
+    def loadCheckpoint(self, path):
+        self.hps = utils.get_hparams_from_file(f"checkpoints/{path}/config.json")
+        self.SVCVITS = SynthesizerTrn(
+            self.hps.data.filter_length // 2 + 1,
+            self.hps.train.segment_size // self.hps.data.hop_length,
+            **self.hps.model)
+        _ = utils.load_checkpoint(f"checkpoints/{path}/model.pdparams", self.SVCVITS, None)
+        _ = self.SVCVITS.eval().to(self.device)
+        self.speakers = self.hps.spk
+
+    def get_units(self, source, sr):
+        source = source.unsqueeze(0).cuda() if self.device == 'gpu:0' else source.unsqueeze(0).cpu()
+        # hubert没有迁移到paddle上。这里也就不迁移了。
+        with torch.inference_mode():
+            units = self.hubert_soft.units(source)
+            return units
+
+
+    def get_unit_pitch(self, in_path, tran):
+        source, sr = torchaudio.load(in_path)
+        source = torchaudio.functional.resample(source, sr, 16000)
+        if len(source.shape) == 2 and source.shape[1] >= 2:
+            source = torch.mean(source, dim=0).unsqueeze(0)
+        soft = self.get_units(source, sr).squeeze(0).cpu().numpy()
+        f0_coarse, f0 = get_f0(source.cpu().numpy()[0], soft.shape[0]*2, tran)
+        return soft, f0
+
+    def infer(self, speaker_id, tran, raw_path):
+        speaker_id = self.speakers[speaker_id]
+        sid = torch.LongTensor([int(speaker_id)]).to(self.device).unsqueeze(0)
+        soft, pitch = self.get_unit_pitch(raw_path, tran)
+        f0 = torch.FloatTensor(clean_pitch(pitch)).unsqueeze(0).to(self.device)
+        stn_tst = torch.FloatTensor(soft)
+        with torch.no_grad():
+            x_tst = stn_tst.unsqueeze(0).to(self.device)
+            x_tst = torch.repeat_interleave(x_tst, repeats=2, dim=1).transpose(1, 2)
+            audio = self.SVCVITS.infer(x_tst, f0=f0, g=sid)[0,0].data.float()
+        return audio, audio.shape[-1]
+
+    def inference(self,srcaudio,chara,tran,slice_db):
+        sampling_rate, audio = srcaudio
+        audio = (audio / np.iinfo(audio.dtype).max).astype(np.float32)
+        if len(audio.shape) > 1:
+            audio = librosa.to_mono(audio.transpose(1, 0))
+        if sampling_rate != 16000:
+            audio = librosa.resample(audio, orig_sr=sampling_rate, target_sr=16000)
+        soundfile.write("tmpwav.wav", audio, 16000, format="wav")
+        chunks = slicer.cut("tmpwav.wav", db_thresh=slice_db)
+        audio_data, audio_sr = slicer.chunks2audio("tmpwav.wav", chunks)
+        audio = []
+        for (slice_tag, data) in audio_data:
+            length = int(np.ceil(len(data) / audio_sr * self.hps.data.sampling_rate))
+            raw_path = io.BytesIO()
+            soundfile.write(raw_path, data, audio_sr, format="wav")
+            raw_path.seek(0)
+            if slice_tag:
+                _audio = np.zeros(length)
+            else:
+                out_audio, out_sr = self.infer(chara, tran, raw_path)
+                _audio = out_audio.cpu().numpy()
+            audio.extend(list(_audio))
+        audio = (np.array(audio) * 32768.0).astype('int16')
+        return (self.hps.data.sampling_rate,audio)

inference/slicer.py

@@ -0,0 +1,142 @@
+import librosa
+import paddle
+import paddle.audio as paddleaudio
+
+
+class Slicer:
+    def __init__(self,
+                 sr: int,
+                 threshold: float = -40.,
+                 min_length: int = 5000,
+                 min_interval: int = 300,
+                 hop_size: int = 20,
+                 max_sil_kept: int = 5000):
+        if not min_length >= min_interval >= hop_size:
+            raise ValueError('The following condition must be satisfied: min_length >= min_interval >= hop_size')
+        if not max_sil_kept >= hop_size:
+            raise ValueError('The following condition must be satisfied: max_sil_kept >= hop_size')
+        min_interval = sr * min_interval / 1000
+        self.threshold = 10 ** (threshold / 20.)
+        self.hop_size = round(sr * hop_size / 1000)
+        self.win_size = min(round(min_interval), 4 * self.hop_size)
+        self.min_length = round(sr * min_length / 1000 / self.hop_size)
+        self.min_interval = round(min_interval / self.hop_size)
+        self.max_sil_kept = round(sr * max_sil_kept / 1000 / self.hop_size)
+
+    def _apply_slice(self, waveform, begin, end):
+        if len(waveform.shape) > 1:
+            return waveform[:, begin * self.hop_size: min(waveform.shape[1], end * self.hop_size)]
+        else:
+            return waveform[begin * self.hop_size: min(waveform.shape[0], end * self.hop_size)]
+
+    # @timeit
+    def slice(self, waveform):
+        if len(waveform.shape) > 1:
+            samples = librosa.to_mono(waveform)
+        else:
+            samples = waveform
+        if samples.shape[0] <= self.min_length:
+            return {"0": {"slice": False, "split_time": f"0,{len(waveform)}"}}
+        rms_list = librosa.feature.rms(y=samples, frame_length=self.win_size, hop_length=self.hop_size).squeeze(0)
+        sil_tags = []
+        silence_start = None
+        clip_start = 0
+        for i, rms in enumerate(rms_list):
+            # Keep looping while frame is silent.
+            if rms < self.threshold:
+                # Record start of silent frames.
+                if silence_start is None:
+                    silence_start = i
+                continue
+            # Keep looping while frame is not silent and silence start has not been recorded.
+            if silence_start is None:
+                continue
+            # Clear recorded silence start if interval is not enough or clip is too short
+            is_leading_silence = silence_start == 0 and i > self.max_sil_kept
+            need_slice_middle = i - silence_start >= self.min_interval and i - clip_start >= self.min_length
+            if not is_leading_silence and not need_slice_middle:
+                silence_start = None
+                continue
+            # Need slicing. Record the range of silent frames to be removed.
+            if i - silence_start <= self.max_sil_kept:
+                pos = rms_list[silence_start: i + 1].argmin() + silence_start
+                if silence_start == 0:
+                    sil_tags.append((0, pos))
+                else:
+                    sil_tags.append((pos, pos))
+                clip_start = pos
+            elif i - silence_start <= self.max_sil_kept * 2:
+                pos = rms_list[i - self.max_sil_kept: silence_start + self.max_sil_kept + 1].argmin()
+                pos += i - self.max_sil_kept
+                pos_l = rms_list[silence_start: silence_start + self.max_sil_kept + 1].argmin() + silence_start
+                pos_r = rms_list[i - self.max_sil_kept: i + 1].argmin() + i - self.max_sil_kept
+                if silence_start == 0:
+                    sil_tags.append((0, pos_r))
+                    clip_start = pos_r
+                else:
+                    sil_tags.append((min(pos_l, pos), max(pos_r, pos)))
+                    clip_start = max(pos_r, pos)
+            else:
+                pos_l = rms_list[silence_start: silence_start + self.max_sil_kept + 1].argmin() + silence_start
+                pos_r = rms_list[i - self.max_sil_kept: i + 1].argmin() + i - self.max_sil_kept
+                if silence_start == 0:
+                    sil_tags.append((0, pos_r))
+                else:
+                    sil_tags.append((pos_l, pos_r))
+                clip_start = pos_r
+            silence_start = None
+        # Deal with trailing silence.
+        total_frames = rms_list.shape[0]
+        if silence_start is not None and total_frames - silence_start >= self.min_interval:
+            silence_end = min(total_frames, silence_start + self.max_sil_kept)
+            pos = rms_list[silence_start: silence_end + 1].argmin() + silence_start
+            sil_tags.append((pos, total_frames + 1))
+        # Apply and return slices.
+        if len(sil_tags) == 0:
+            return {"0": {"slice": False, "split_time": f"0,{len(waveform)}"}}
+        else:
+            chunks = []
+            # 第一段静音并非从头开始，补上有声片段
+            if sil_tags[0][0]:
+                chunks.append(
+                    {"slice": False, "split_time": f"0,{min(waveform.shape[0], sil_tags[0][0] * self.hop_size)}"})
+            for i in range(0, len(sil_tags)):
+                # 标识有声片段（跳过第一段）
+                if i:
+                    chunks.append({"slice": False,
+                                   "split_time": f"{sil_tags[i - 1][1] * self.hop_size},{min(waveform.shape[0], sil_tags[i][0] * self.hop_size)}"})
+                # 标识所有静音片段
+                chunks.append({"slice": True,
+                               "split_time": f"{sil_tags[i][0] * self.hop_size},{min(waveform.shape[0], sil_tags[i][1] * self.hop_size)}"})
+            # 最后一段静音并非结尾，补上结尾片段
+            if sil_tags[-1][1] * self.hop_size < len(waveform):
+                chunks.append({"slice": False, "split_time": f"{sil_tags[-1][1] * self.hop_size},{len(waveform)}"})
+            chunk_dict = {}
+            for i in range(len(chunks)):
+                chunk_dict[str(i)] = chunks[i]
+            return chunk_dict
+
+
+def cut(audio_path, db_thresh=-30, min_len=5000):
+    audio, sr = librosa.load(audio_path, sr=None)
+    slicer = Slicer(
+        sr=sr,
+        threshold=db_thresh,
+        min_length=min_len
+    )
+    chunks = slicer.slice(audio)
+    return chunks
+
+
+def chunks2audio(audio_path, chunks):
+    chunks = dict(chunks)
+    audio, sr = paddleaudio.load(audio_path)
+    if len(audio.shape) == 2 and audio.shape[1] >= 2:
+        audio = paddle.mean(audio, axis=0).unsqueeze(0)
+    audio = audio.cpu().numpy()[0]
+    result = []
+    for k, v in chunks.items():
+        tag = v["split_time"].split(",")
+        if tag[0] != tag[1]:
+            result.append((v["slice"], audio[int(tag[0]):int(tag[1])]))
+    return result, sr

inference_main.py

@@ -0,0 +1,108 @@
+import io
+import logging
+import time
+from pathlib import Path
+
+import librosa
+import matplotlib.pyplot as plt
+import numpy as np
+import soundfile
+
+from inference import infer_tool
+from inference import slicer
+from inference.infer_tool import Svc
+
+logging.getLogger('numba').setLevel(logging.WARNING)
+chunks_dict = infer_tool.read_temp("inference/chunks_temp.json")
+
+# 这里是推理用到的所有参数，从这里修改参数即可
+模型路径:str = "./logs/44k/G_10000.pdparams"            # 模型路径
+推理文件列表:list = ["1.wav"] # wav文件名列表，放在raw文件夹下
+音高调整:list = [0]                                  # 音高调整，支持正负（半音）
+合成目标说话人名称:list = ['yuuka']                     # 合成目标说话人名称
+自动预测音高:bool = False                             # 语音转换自动预测音高，转换歌声时不要打开这个会严重跑调
+聚类模型路径:str = "logs/44k/kmeans_10000.pdparams"   # 聚类模型路径，如果没有训练聚类则随便填
+聚类方案占比:float = 0                                # 聚类方案占比，范围0-1，若没有训练聚类模型则填0即可
+静音分贝:int = -40                                    # 静音分贝阈值，默认-40，嘈杂的音频可以-30，干声保留呼吸可以-50
+推理设备:str or None = None                           # 推理设备，None则为自动选择cpu和gpu
+音频输出格式:str = 'flac'                             # 音频输出格式
+噪音比例:float = 0.4                                  # 声音有点电的话可以尝试调高这个，但是会降低音质，较为玄学
+
+def main():
+    import argparse
+
+    parser = argparse.ArgumentParser(description='飞桨sovits4 推理模块')
+    parser.add_argument('-m', '--model_path', type=str, default=模型路径, help='模型路径')
+    parser.add_argument('-c', '--config_path', type=str, default="./logs/44k/config.json", help='配置文件路径')
+    parser.add_argument('-n', '--clean_names', type=str, nargs='+', default=推理文件列表, help='wav文件名列表，放在raw文件夹下')
+    parser.add_argument('-t', '--trans', type=int, nargs='+', default=音高调整, help='音高调整，支持正负（半音）')
+    parser.add_argument('-s', '--spk_list', type=str, nargs='+', default=合成目标说话人名称, help='合成目标说话人名称')
+
+    parser.add_argument('-a', '--auto_predict_f0', action='store_true', default=自动预测音高,
+                        help='语音转换自动预测音高，转换歌声时不要打开这个会严重跑调')
+    parser.add_argument('-cm', '--cluster_model_path', type=str, default=聚类模型路径, help='聚类模型路径，如果没有训练聚类则随便填')
+    parser.add_argument('-cr', '--cluster_infer_ratio', type=float, default=聚类方案占比, help='聚类方案占比，范围0-1，若没有训练聚类模型则填0即可')
+
+    parser.add_argument('-sd', '--slice_db', type=int, default=静音分贝, help='默认-40，嘈杂的音频可以-30，干声保留呼吸可以-50')
+    parser.add_argument('-d', '--device', type=str, default=推理设备, help='推理设备，None则为自动选择cpu和gpu')
+    parser.add_argument('-ns', '--noice_scale', type=float, default=噪音比例, help='噪音级别，会影响咬字和音质，较为玄学')
+    parser.add_argument('-p', '--pad_seconds', type=float, default=0.5, help='推理音频pad秒数，由于未知原因开头结尾会有异响，pad一小段静音段后就不会出现')
+    parser.add_argument('-wf', '--wav_format', type=str, default=音频输出格式, help='音频输出格式')
+
+    args = parser.parse_args()
+
+    svc_model = Svc(args.model_path, args.config_path, args.device, args.cluster_model_path)
+    infer_tool.mkdir(["raw", "results"])
+    clean_names = args.clean_names
+    trans = args.trans
+    spk_list = args.spk_list
+    slice_db = args.slice_db
+    wav_format = args.wav_format
+    auto_predict_f0 = args.auto_predict_f0
+    cluster_infer_ratio = args.cluster_infer_ratio
+    noice_scale = args.noice_scale
+    pad_seconds = args.pad_seconds
+
+    infer_tool.fill_a_to_b(trans, clean_names)
+    for clean_name, tran in zip(clean_names, trans):
+        raw_audio_path = f"raw/{clean_name}"
+        if "." not in raw_audio_path:
+            raw_audio_path += ".wav"
+        infer_tool.format_wav(raw_audio_path)
+        wav_path = Path(raw_audio_path).with_suffix('.wav')
+        chunks = slicer.cut(wav_path, db_thresh=slice_db)
+        audio_data, audio_sr = slicer.chunks2audio(wav_path, chunks)
+
+        for spk in spk_list:
+            audio = []
+            for (slice_tag, data) in audio_data:
+                print(f'#=====分段开始，{round(len(data) / audio_sr, 3)}秒======')
+
+                length = int(np.ceil(len(data) / audio_sr * svc_model.target_sample))
+                if slice_tag:
+                    print('跳过空段')
+                    _audio = np.zeros(length)
+                else:
+                    # padd
+                    pad_len = int(audio_sr * pad_seconds)
+                    data = np.concatenate([np.zeros([pad_len]), data, np.zeros([pad_len])])
+                    raw_path = io.BytesIO()
+                    soundfile.write(raw_path, data, audio_sr, format="wav")
+                    raw_path.seek(0)
+                    out_audio, out_sr = svc_model.infer(spk, tran, raw_path,
+                                                        cluster_infer_ratio=cluster_infer_ratio,
+                                                        auto_predict_f0=auto_predict_f0,
+                                                        noice_scale=noice_scale
+                                                        )
+                    _audio = out_audio.detach().cpu().numpy()
+                    pad_len = int(svc_model.target_sample * pad_seconds)
+                    _audio = _audio[pad_len:-pad_len]
+
+                audio.extend(list(infer_tool.pad_array(_audio, length)))
+            key = "auto" if auto_predict_f0 else f"{tran}key"
+            cluster_name = "" if cluster_infer_ratio == 0 else f"_{cluster_infer_ratio}"
+            res_path = f'./results/{clean_name}_{key}_{spk}{cluster_name}.{wav_format}'
+            soundfile.write(res_path, audio, svc_model.target_sample, format=wav_format)
+
+if __name__ == '__main__':
+    main()

logs/44k/config.json

@@ -0,0 +1,95 @@
+{
+    "train": {
+        "log_interval": 800,
+        "eval_interval": 400,
+        "seed": 1234,
+        "epochs": 114514,
+        "learning_rate": 0.0001,
+        "betas": [
+            0.8,
+            0.99
+        ],
+        "eps": 1e-05,
+        "batch_size": 2,
+        "fp16_run": true,
+        "lr_decay": 0.999875,
+        "segment_size": 10240,
+        "init_lr_ratio": 1,
+        "warmup_epochs": 0,
+        "c_mel": 45,
+        "c_kl": 1.0,
+        "use_sr": true,
+        "max_speclen": 512,
+        "port": "8001",
+        "keep_ckpts": 5
+    },
+    "data": {
+        "training_files": "filelists/train.txt",
+        "validation_files": "filelists/val.txt",
+        "max_wav_value": 32768.0,
+        "sampling_rate": 44100,
+        "filter_length": 2048,
+        "hop_length": 512,
+        "win_length": 2048,
+        "n_mel_channels": 80,
+        "mel_fmin": 0.0,
+        "mel_fmax": 22050
+    },
+    "model": {
+        "inter_channels": 192,
+        "hidden_channels": 192,
+        "filter_channels": 768,
+        "n_heads": 2,
+        "n_layers": 6,
+        "kernel_size": 3,
+        "p_dropout": 0.1,
+        "resblock": "1",
+        "resblock_kernel_sizes": [
+            3,
+            7,
+            11
+        ],
+        "resblock_dilation_sizes": [
+            [
+                1,
+                3,
+                5
+            ],
+            [
+                1,
+                3,
+                5
+            ],
+            [
+                1,
+                3,
+                5
+            ]
+        ],
+        "upsample_rates": [
+            8,
+            8,
+            2,
+            2,
+            2
+        ],
+        "upsample_initial_channel": 512,
+        "upsample_kernel_sizes": [
+            16,
+            16,
+            4,
+            4,
+            4
+        ],
+        "n_layers_q": 3,
+        "use_spectral_norm": false,
+        "gin_channels": 256,
+        "ssl_dim": 256,
+        "n_speakers": 200
+    },
+    "spk": {
+        "yuuka": 0
+    },
+    "clean_logs": true,
+    "trainer": "admin"
+}

logs/44k/train.log

@@ -0,0 +1,8 @@
+2023-07-22 21:55:56,616	44k	INFO	{'train': {'log_interval': 800, 'eval_interval': 400, 'seed': 1234, 'epochs': 114514, 'learning_rate': 0.0001, 'betas': [0.8, 0.99], 'eps': 1e-05, 'batch_size': 2, 'fp16_run': True, 'lr_decay': 0.999875, 'segment_size': 10240, 'init_lr_ratio': 1, 'warmup_epochs': 0, 'c_mel': 45, 'c_kl': 1.0, 'use_sr': True, 'max_speclen': 512, 'port': '8001', 'keep_ckpts': 5}, 'data': {'training_files': 'filelists/train.txt', 'validation_files': 'filelists/val.txt', 'max_wav_value': 32768.0, 'sampling_rate': 44100, 'filter_length': 2048, 'hop_length': 512, 'win_length': 2048, 'n_mel_channels': 80, 'mel_fmin': 0.0, 'mel_fmax': 22050}, 'model': {'inter_channels': 192, 'hidden_channels': 192, 'filter_channels': 768, 'n_heads': 2, 'n_layers': 6, 'kernel_size': 3, 'p_dropout': 0.1, 'resblock': '1', 'resblock_kernel_sizes': [3, 7, 11], 'resblock_dilation_sizes': [[1, 3, 5], [1, 3, 5], [1, 3, 5]], 'upsample_rates': [8, 8, 2, 2, 2], 'upsample_initial_channel': 512, 'upsample_kernel_sizes': [16, 16, 4, 4, 4], 'n_layers_q': 3, 'use_spectral_norm': False, 'gin_channels': 256, 'ssl_dim': 256, 'n_speakers': 200}, 'spk': {'yuuka': 0}, 'clean_logs': True, 'trainer': 'admin', 'model_dir': './logs/44k'}
+2023-07-22 21:55:56,617	44k	WARNING	/home/aistudio/build不是git存储库，因此将忽略哈希值比较。
+2023-07-22 21:55:59,336	44k	INFO	加载检查点 './logs/44k/G_0.pdparams' (迭代次数 1)
+2023-07-22 21:55:59,680	44k	INFO	加载检查点 './logs/44k/D_0.pdparams' (迭代次数 1)
+2023-07-22 21:56:12,355	44k	INFO	训练回合：1 [0%]
+2023-07-22 21:56:12,356	44k	INFO	损失：[2.723755359649658, 2.7280983924865723, 7.272645950317383, 30.232248306274414, 3.609935998916626]，步数：0，学习率：0.0001
+2023-07-22 21:56:20,694	44k	INFO	保存模型和优化器状态位于迭代次数1 到 ./logs/44k/G_0.pdparams
+2023-07-22 21:56:22,248	44k	INFO	保存模型和优化器状态位于迭代次数1 到 ./logs/44k/D_0.pdparams

models.py

@@ -0,0 +1,556 @@
+import copy
+import math
+import paddle
+from paddle import nn
+from paddle.nn import functional as F
+
+import modules.attentions as attentions
+import modules.commons as commons
+import modules.modules as modules
+
+from paddle.nn import Conv1D, Conv1DTranspose, AvgPool1D, Conv2D
+from paddle.nn.utils import weight_norm, remove_weight_norm, spectral_norm
+
+import utils
+from modules.commons import init_weights, get_padding
+from vdecoder.hifigan.models import Generator
+from utils import f0_to_coarse
+import random
+import string
+import time
+
+class ResidualCouplingBlock(nn.Layer):
+  def __init__(self,
+      channels,
+      hidden_channels,
+      kernel_size,
+      dilation_rate,
+      n_layers,
+      n_flows=4,
+      gin_channels=0):
+    super().__init__()
+    self.channels = channels
+    self.hidden_channels = hidden_channels
+    self.kernel_size = kernel_size
+    self.dilation_rate = dilation_rate
+    self.n_layers = n_layers
+    self.n_flows = n_flows
+    self.gin_channels = gin_channels
+
+    self.flows = nn.LayerList()
+    for i in range(n_flows):
+      self.flows.append(modules.ResidualCouplingLayer(channels, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels, mean_only=True))
+      self.flows.append(modules.Flip())
+
+  def forward(self, x, x_mask, g=None, reverse=False):
+    if not reverse:
+      for flow in self.flows:
+        x, _ = flow(x, x_mask, g=g, reverse=reverse)
+    else:
+      for flow in reversed(self.flows):
+        x = flow(x, x_mask, g=g, reverse=reverse)
+    return x
+
+
+class Encoder(nn.Layer):
+  def __init__(self,
+      in_channels,
+      out_channels,
+      hidden_channels,
+      kernel_size,
+      dilation_rate,
+      n_layers,
+      gin_channels=0):
+    super().__init__()
+    self.in_channels = in_channels
+    self.out_channels = out_channels
+    self.hidden_channels = hidden_channels
+    self.kernel_size = kernel_size
+    self.dilation_rate = dilation_rate
+    self.n_layers = n_layers
+    self.gin_channels = gin_channels
+
+    self.pre = nn.Conv1D(in_channels, hidden_channels, 1)
+    self.enc = modules.WN(hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels)
+    self.proj = nn.Conv1D(hidden_channels, out_channels * 2, 1)
+
+  def forward(self, x, x_lengths, g=None):
+    # print(x.shape,x_lengths.shape)
+    x_mask = paddle.unsqueeze(commons.sequence_mask(x_lengths, x.shape[2]), 1).cast(x.dtype)
+    x = self.pre(x) * x_mask
+    x = self.enc(x, x_mask, g=g)
+    stats = self.proj(x) * x_mask
+    m, logs = paddle.split(stats, [self.out_channels,self.out_channels], axis=1)
+    z = (m + paddle.randn(m.shape,m.dtype) * paddle.exp(logs)) * x_mask
+    return z, m, logs, x_mask
+
+
+class TextEncoder(nn.Layer):
+  def __init__(self,
+      out_channels,
+      hidden_channels,
+      kernel_size,
+      n_layers,
+      gin_channels=0,
+      filter_channels=None,
+      n_heads=None,
+      p_dropout=None):
+    super().__init__()
+    self.out_channels = out_channels
+    self.hidden_channels = hidden_channels
+    self.kernel_size = kernel_size
+    self.n_layers = n_layers
+    self.gin_channels = gin_channels
+    self.proj = nn.Conv1D(hidden_channels, out_channels * 2, 1)
+    self.f0_emb = nn.Embedding(256, hidden_channels)
+
+    self.enc_ =  attentions.Encoder(
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout)
+
+  def forward(self, x, x_mask, f0=None, noice_scale=1):
+    x = x + self.f0_emb(f0).transpose((0,2,1))
+    x = self.enc_(x * x_mask, x_mask)
+    stats = self.proj(x) * x_mask
+    m, logs = paddle.split(stats, [self.out_channels,self.out_channels], axis = 1)
+    z = (m + paddle.randn(m.shape,m.dtype) * paddle.exp(logs) * noice_scale) * x_mask
+    return z, m, logs, x_mask
+
+
+
+class DiscriminatorP(paddle.nn.Layer):
+    def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
+        super(DiscriminatorP, self).__init__()
+        self.period = period
+        self.use_spectral_norm = use_spectral_norm
+        norm_f = weight_norm if use_spectral_norm == False else spectral_norm
+        self.convs = nn.LayerList([
+            norm_f(Conv2D(1, 32, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
+            norm_f(Conv2D(32, 128, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
+            norm_f(Conv2D(128, 512, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
+            norm_f(Conv2D(512, 1024, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
+            norm_f(Conv2D(1024, 1024, (kernel_size, 1), 1, padding=(get_padding(kernel_size, 1), 0))),
+        ])
+        self.conv_post = norm_f(Conv2D(1024, 1, (3, 1), 1, padding=(1, 0)))
+
+    def forward(self, x):
+        fmap = []
+
+        # 1d to 2d
+        b, c, t = x.shape
+        if t % self.period != 0: # pad first
+            n_pad = self.period - (t % self.period)
+            x = F.pad(x, (0, n_pad), "reflect",data_format='NCL')
+            t = t + n_pad
+        x = x.reshape((b, c, t // self.period, self.period))
+
+        for l in self.convs:
+            x = l(x)
+            x = F.leaky_relu(x, modules.LRELU_SLOPE)
+            fmap.append(x)
+        x = self.conv_post(x)
+        fmap.append(x)
+        x = paddle.flatten(x, 1, -1)
+
+        return x, fmap
+
+
+class DiscriminatorS(paddle.nn.Layer):
+    def __init__(self, use_spectral_norm=False):
+        super(DiscriminatorS, self).__init__()
+        norm_f = weight_norm if use_spectral_norm == False else spectral_norm
+        self.convs = nn.LayerList([
+            norm_f(Conv1D(1, 16, 15, 1, padding=7)),
+            norm_f(Conv1D(16, 64, 41, 4, groups=4, padding=20)),
+            norm_f(Conv1D(64, 256, 41, 4, groups=16, padding=20)),
+            norm_f(Conv1D(256, 1024, 41, 4, groups=64, padding=20)),
+            norm_f(Conv1D(1024, 1024, 41, 4, groups=256, padding=20)),
+            norm_f(Conv1D(1024, 1024, 5, 1, padding=2)),
+        ])
+        self.conv_post = norm_f(Conv1D(1024, 1, 3, 1, padding=1))
+
+    def forward(self, x):
+        fmap = []
+
+        for l in self.convs:
+            x = l(x)
+            x = F.leaky_relu(x, modules.LRELU_SLOPE)
+            fmap.append(x)
+        x = self.conv_post(x)
+        fmap.append(x)
+        x = paddle.flatten(x, 1, -1)
+
+        return x, fmap
+
+
+class MultiPeriodDiscriminator(paddle.nn.Layer):
+    def __init__(self, use_spectral_norm=False):
+        super(MultiPeriodDiscriminator, self).__init__()
+        periods = [2,3,5,7,11]
+
+        discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
+        discs = discs + [DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods]
+        self.discriminators = nn.LayerList(discs)
+
+    def forward(self, y, y_hat):
+        y_d_rs = []
+        y_d_gs = []
+        fmap_rs = []
+        fmap_gs = []
+        for i, d in enumerate(self.discriminators):
+            y_d_r, fmap_r = d(y)
+            y_d_g, fmap_g = d(y_hat)
+            y_d_rs.append(y_d_r)
+            y_d_gs.append(y_d_g)
+            fmap_rs.append(fmap_r)
+            fmap_gs.append(fmap_g)
+
+        return y_d_rs, y_d_gs, fmap_rs, fmap_gs
+
+
+class SpeakerEncoder(paddle.nn.Layer):
+    def __init__(self, mel_n_channels=80, model_num_layers=3, model_hidden_size=256, model_embedding_size=256):
+        super(SpeakerEncoder, self).__init__()
+        self.lstm = nn.LSTM(mel_n_channels, model_hidden_size, model_num_layers)
+        self.linear = nn.Linear(model_hidden_size, model_embedding_size)
+        self.relu = nn.ReLU()
+
+    def forward(self, mels):
+        self.lstm.flatten_parameters()
+        _, (hidden, _) = self.lstm(mels)
+        embeds_raw = self.relu(self.linear(hidden[-1]))
+        return embeds_raw / paddle.norm(embeds_raw, axis=1, keepdim=True)
+
+    def compute_partial_slices(self, total_frames, partial_frames, partial_hop):
+        mel_slices = []
+        for i in range(0, total_frames-partial_frames, partial_hop):
+            mel_range = paddle.arange(i, i+partial_frames)
+            mel_slices.append(mel_range)
+
+        return mel_slices
+
+    def embed_utterance(self, mel, partial_frames=128, partial_hop=64):
+        mel_len = mel.shape[1]
+        last_mel = mel[:,-partial_frames:]
+
+        if mel_len > partial_frames:
+            mel_slices = self.compute_partial_slices(mel_len, partial_frames, partial_hop)
+            mels = list(mel[:,s] for s in mel_slices)
+            mels.append(last_mel)
+            mels = paddle.stack(tuple(mels), 0).squeeze(1)
+
+            with paddle.no_grad():
+                partial_embeds = self(mels)
+            embed = paddle.mean(partial_embeds, axis=0).unsqueeze(0)
+            #embed = embed / torch.linalg.norm(embed, 2)
+        else:
+            with paddle.no_grad():
+                embed = self(last_mel)
+
+        return embed
+
+class F0Decoder(nn.Layer):
+    def __init__(self,
+                 out_channels,
+                 hidden_channels,
+                 filter_channels,
+                 n_heads,
+                 n_layers,
+                 kernel_size,
+                 p_dropout,
+                 spk_channels=0):
+        super().__init__()
+        self.out_channels = out_channels
+        self.hidden_channels = hidden_channels
+        self.filter_channels = filter_channels
+        self.n_heads = n_heads
+        self.n_layers = n_layers
+        self.kernel_size = kernel_size
+        self.p_dropout = p_dropout
+        self.spk_channels = spk_channels
+
+        self.prenet = nn.Conv1D(hidden_channels, hidden_channels, 3, padding=1)
+        self.decoder = attentions.FFT(
+            hidden_channels,
+            filter_channels,
+            n_heads,
+            n_layers,
+            kernel_size,
+            p_dropout)
+        self.proj = nn.Conv1D(hidden_channels, out_channels, 1)
+        self.f0_prenet = nn.Conv1D(1, hidden_channels , 3, padding=1)
+        self.cond = nn.Conv1D(spk_channels, hidden_channels, 1)
+
+    def forward(self, x, norm_f0, x_mask, spk_emb=None):
+        x = x.detach()
+        if (spk_emb is not None):
+            x = x + self.cond(spk_emb)
+        x += self.f0_prenet(norm_f0)
+        x = self.prenet(x) * x_mask
+        x = self.decoder(x * x_mask, x_mask)
+        x = self.proj(x) * x_mask
+        return x
+
+class SynthesizerTrn_test(nn.Layer):
+  """
+  Synthesizer for Training
+  """
+
+  def __init__(self,
+    spec_channels,
+    segment_size,
+    inter_channels,
+    hidden_channels,
+    filter_channels,
+    n_heads,
+    n_layers,
+    kernel_size,
+    p_dropout,
+    resblock,
+    resblock_kernel_sizes,
+    resblock_dilation_sizes,
+    upsample_rates,
+    upsample_initial_channel,
+    upsample_kernel_sizes,
+    gin_channels,
+    ssl_dim,
+    n_speakers,
+    sampling_rate=44100,
+    **kwargs):
+
+    super().__init__()
+    self.spec_channels = spec_channels
+    self.inter_channels = inter_channels
+    self.hidden_channels = hidden_channels
+    self.filter_channels = filter_channels
+    self.n_heads = n_heads
+    self.n_layers = n_layers
+    self.kernel_size = kernel_size
+    self.p_dropout = p_dropout
+    self.resblock = resblock
+    self.resblock_kernel_sizes = resblock_kernel_sizes
+    self.resblock_dilation_sizes = resblock_dilation_sizes
+    self.upsample_rates = upsample_rates
+    self.upsample_initial_channel = upsample_initial_channel
+    self.upsample_kernel_sizes = upsample_kernel_sizes
+    self.segment_size = segment_size
+    self.gin_channels = gin_channels
+    self.ssl_dim = ssl_dim
+
+    init = paddle.nn.initializer.Normal(0.001,1)
+    pa = paddle.ParamAttr(f'emb_g_pa_{int(time.time())}',init)
+    self.emb_g = nn.Embedding(n_speakers, gin_channels, weight_attr = pa)
+
+    init = paddle.nn.initializer.Normal(2.7973e-06,0.0161)
+    pre_pa = paddle.ParamAttr(f'pre_pa_{int(time.time())}',init)
+    self.pre = nn.Conv1D(ssl_dim, hidden_channels, kernel_size=5, padding=2, weight_attr = pre_pa)
+
+    self.enc_p = TextEncoder(
+        inter_channels,
+        hidden_channels,
+        filter_channels=filter_channels,
+        n_heads=n_heads,
+        n_layers=n_layers,
+        kernel_size=kernel_size,
+        p_dropout=p_dropout
+    )
+    hps = {
+        "sampling_rate": sampling_rate,
+        "inter_channels": inter_channels,
+        "resblock": resblock,
+        "resblock_kernel_sizes": resblock_kernel_sizes,
+        "resblock_dilation_sizes": resblock_dilation_sizes,
+        "upsample_rates": upsample_rates,
+        "upsample_initial_channel": upsample_initial_channel,
+        "upsample_kernel_sizes": upsample_kernel_sizes,
+        "gin_channels": gin_channels,
+    }
+    self.dec = Generator(h=hps)
+    self.enc_q = Encoder(spec_channels, inter_channels, hidden_channels, 5, 1, 16, gin_channels=gin_channels)
+    self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, 4, gin_channels=gin_channels)
+    self.f0_decoder = F0Decoder(
+        1,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout,
+        spk_channels=gin_channels
+    )
+    initer = paddle.nn.initializer.Normal(mean = 0.202, std = 0.9640, name = f'emb_uv_init_weight_{time.time}')
+    emb_uv_pa = paddle.ParamAttr(f'emb_uv_init_weight_pa_{int(time.time())}',initer)
+    self.emb_uv = nn.Embedding(2, hidden_channels, weight_attr = emb_uv_pa)
+
+  def forward(self, c, f0, uv, spec, g=None, c_lengths=None, spec_lengths=None):
+    g = self.emb_g(g).transpose([0,2,1])
+    # ssl prenet
+    x_mask = paddle.unsqueeze(commons.sequence_mask(c_lengths, c.shape[2]), 1).astype(c.dtype)
+    emb_uv = self.emb_uv(uv.cast('int64')).transpose([0,2,1])
+    prec = self.pre(c)
+    x = prec * x_mask + emb_uv
+    # f0 predict
+    lf0 = 2595. * paddle.log10(1. + f0.unsqueeze(1) / 700.) / 500
+    norm_lf0 = utils.normalize_f0(lf0, x_mask, uv)
+    pred_lf0 = self.f0_decoder(x, norm_lf0, x_mask, spk_emb=g)
+
+    # encoder
+    z_ptemp, m_p, logs_p, _ = self.enc_p(x, x_mask, f0 = f0_to_coarse(f0))
+    z, m_q, logs_q, spec_mask = self.enc_q(spec, spec_lengths, g=g) 
+
+    # flow
+    z_p = self.flow(z, spec_mask, g=g)
+    z_slice, pitch_slice, ids_slice = commons.rand_slice_segments_with_pitch(z, f0, spec_lengths, self.segment_size)
+    # nsf decoder
+    o = self.dec(z_slice, g=g, f0=pitch_slice)
+
+    return o, ids_slice, spec_mask, (z, z_p, m_p, logs_p, m_q, logs_q), pred_lf0, norm_lf0, lf0
+
+  def infer(self, c, f0, uv, g=None, noice_scale=0.35, predict_f0=False):
+    c_lengths = ((paddle.ones((c.shape[0],)) * c.shape[-1])).cpu() if 'cpu'in str(c.place) else ((paddle.ones((c.shape[0],)) * c.shape[-1])).cuda()
+    g = self.emb_g(g).transpose([0,2,1])
+    x_mask = paddle.unsqueeze(commons.sequence_mask(c_lengths, c.shape[2]), 1).astype(c.dtype)
+    x = self.pre(c) * x_mask + self.emb_uv(uv.astype('int64')).transpose([0,2,1])
+    if predict_f0:
+        lf0 = 2595. * paddle.log10(1. + f0.unsqueeze(1) / 700.) / 500
+        norm_lf0 = utils.normalize_f0(lf0, x_mask, uv, random_scale=False)
+        pred_lf0 = self.f0_decoder(x, norm_lf0, x_mask, spk_emb=g)
+        f0 = (700 * (paddle.pow(paddle.to_tensor(10.), pred_lf0 * 500 / 2595) - 1)).squeeze(1)
+    z_p, m_p, logs_p, c_mask = self.enc_p(x, x_mask, f0=f0_to_coarse(f0), noice_scale=noice_scale)
+    z = self.flow(z_p, c_mask, g=g, reverse=True)
+    o = self.dec(z * c_mask, g=g, f0=f0)
+    return o
+
+class SynthesizerTrn(nn.Layer):
+  """
+  Synthesizer for Training
+  """
+
+  def __init__(self,
+    spec_channels,
+    segment_size,
+    inter_channels,
+    hidden_channels,
+    filter_channels,
+    n_heads,
+    n_layers,
+    kernel_size,
+    p_dropout,
+    resblock,
+    resblock_kernel_sizes,
+    resblock_dilation_sizes,
+    upsample_rates,
+    upsample_initial_channel,
+    upsample_kernel_sizes,
+    gin_channels,
+    ssl_dim,
+    n_speakers,
+    sampling_rate=44100,
+    **kwargs):
+
+    super().__init__()
+    self.spec_channels = spec_channels
+    self.inter_channels = inter_channels
+    self.hidden_channels = hidden_channels
+    self.filter_channels = filter_channels
+    self.n_heads = n_heads
+    self.n_layers = n_layers
+    self.kernel_size = kernel_size
+    self.p_dropout = p_dropout
+    self.resblock = resblock
+    self.resblock_kernel_sizes = resblock_kernel_sizes
+    self.resblock_dilation_sizes = resblock_dilation_sizes
+    self.upsample_rates = upsample_rates
+    self.upsample_initial_channel = upsample_initial_channel
+    self.upsample_kernel_sizes = upsample_kernel_sizes
+    self.segment_size = segment_size
+    self.gin_channels = gin_channels
+    self.ssl_dim = ssl_dim
+
+    init = paddle.nn.initializer.Normal(0.001,1)
+    pa = paddle.ParamAttr('emb_g_pa',init)
+    self.emb_g = nn.Embedding(n_speakers, gin_channels, weight_attr = pa)
+
+    init = paddle.nn.initializer.Normal(2.7973e-06,0.0161)
+    pre_pa = paddle.ParamAttr('pre_pa',init)
+    self.pre = nn.Conv1D(ssl_dim, hidden_channels, kernel_size=5, padding=2, weight_attr = pre_pa)
+
+    self.enc_p = TextEncoder(
+        inter_channels,
+        hidden_channels,
+        filter_channels=filter_channels,
+        n_heads=n_heads,
+        n_layers=n_layers,
+        kernel_size=kernel_size,
+        p_dropout=p_dropout
+    )
+    hps = {
+        "sampling_rate": sampling_rate,
+        "inter_channels": inter_channels,
+        "resblock": resblock,
+        "resblock_kernel_sizes": resblock_kernel_sizes,
+        "resblock_dilation_sizes": resblock_dilation_sizes,
+        "upsample_rates": upsample_rates,
+        "upsample_initial_channel": upsample_initial_channel,
+        "upsample_kernel_sizes": upsample_kernel_sizes,
+        "gin_channels": gin_channels,
+    }
+    self.dec = Generator(h=hps)
+    self.enc_q = Encoder(spec_channels, inter_channels, hidden_channels, 5, 1, 16, gin_channels=gin_channels)
+    self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, 4, gin_channels=gin_channels)
+    self.f0_decoder = F0Decoder(
+        1,
+        hidden_channels,
+        filter_channels,
+        n_heads,
+        n_layers,
+        kernel_size,
+        p_dropout,
+        spk_channels=gin_channels
+    )
+    initer = paddle.nn.initializer.Normal(mean = 0.202, std = 0.9640, name = f'emb_uv_init_weight')
+    emb_uv_pa = paddle.ParamAttr('emb_uv_init_weight_pa',initer)
+    self.emb_uv = nn.Embedding(2, hidden_channels, weight_attr = emb_uv_pa)
+
+  def forward(self, c, f0, uv, spec, g=None, c_lengths=None, spec_lengths=None):
+    g = self.emb_g(g).transpose([0,2,1])
+    # ssl prenet
+    x_mask = paddle.unsqueeze(commons.sequence_mask(c_lengths, c.shape[2]), 1).astype(c.dtype)
+    emb_uv = self.emb_uv(uv.cast('int64')).transpose([0,2,1])
+    prec = self.pre(c)
+    x = prec * x_mask + emb_uv
+    # f0 predict
+    lf0 = 2595. * paddle.log10(1. + f0.unsqueeze(1) / 700.) / 500
+    norm_lf0 = utils.normalize_f0(lf0, x_mask, uv)
+    pred_lf0 = self.f0_decoder(x, norm_lf0, x_mask, spk_emb=g)
+
+    # encoder
+    z_ptemp, m_p, logs_p, _ = self.enc_p(x, x_mask, f0 = f0_to_coarse(f0))
+    z, m_q, logs_q, spec_mask = self.enc_q(spec, spec_lengths, g=g) 
+
+    # flow
+    z_p = self.flow(z, spec_mask, g=g)
+    z_slice, pitch_slice, ids_slice = commons.rand_slice_segments_with_pitch(z, f0, spec_lengths, self.segment_size)
+    # nsf decoder
+    o = self.dec(z_slice, g=g, f0=pitch_slice)
+
+    return o, ids_slice, spec_mask, (z, z_p, m_p, logs_p, m_q, logs_q), pred_lf0, norm_lf0, lf0
+
+  def infer(self, c, f0, uv, g=None, noice_scale=0.35, predict_f0=False):
+    c_lengths = ((paddle.ones((c.shape[0],)) * c.shape[-1])).cpu() if 'cpu'in str(c.place) else ((paddle.ones((c.shape[0],)) * c.shape[-1])).cuda()
+    g = self.emb_g(g).transpose([0,2,1])
+    x_mask = paddle.unsqueeze(commons.sequence_mask(c_lengths, c.shape[2]), 1).astype(c.dtype)
+    x = self.pre(c) * x_mask + self.emb_uv(uv.astype('int64')).transpose([0,2,1])
+    if predict_f0:
+        lf0 = 2595. * paddle.log10(1. + f0.unsqueeze(1) / 700.) / 500
+        norm_lf0 = utils.normalize_f0(lf0, x_mask, uv, random_scale=False)
+        pred_lf0 = self.f0_decoder(x, norm_lf0, x_mask, spk_emb=g)
+        f0 = (700 * (paddle.pow(paddle.to_tensor(10.), pred_lf0 * 500 / 2595) - 1)).squeeze(1)
+    z_p, m_p, logs_p, c_mask = self.enc_p(x, x_mask, f0=f0_to_coarse(f0), noice_scale=noice_scale)
+    z = self.flow(z_p, c_mask, g=g, reverse=True)
+    o = self.dec(z * c_mask, g=g, f0=f0)
+    return o

modules/__init__.py

@@ -0,0 +1 @@
+'''由梅花三弄再回首花了一个下午迁移的模块。'''

modules/attentions.py

@@ -0,0 +1,377 @@
+import copy
+import math
+import numpy as np
+import paddle
+from paddle import nn
+from paddle.nn import functional as F
+
+import modules.commons as commons
+import modules.modules as modules
+from modules.modules import LayerNorm
+
+
+class FFT(nn.Layer):
+  def __init__(self, hidden_channels, filter_channels, n_heads, n_layers=1, kernel_size=1, p_dropout=0.,
+               proximal_bias=False, proximal_init=True, **kwargs):
+    super().__init__()
+    self.hidden_channels = hidden_channels
+    self.filter_channels = filter_channels
+    self.n_heads = n_heads
+    self.n_layers = n_layers
+    self.kernel_size = kernel_size
+    self.p_dropout = p_dropout
+    self.proximal_bias = proximal_bias
+    self.proximal_init = proximal_init
+
+    self.drop = nn.Dropout(p_dropout)
+    self.self_attn_layers = nn.LayerList()
+    self.norm_layers_0 = nn.LayerList()
+    self.ffn_layers = nn.LayerList()
+    self.norm_layers_1 = nn.LayerList()
+    for i in range(self.n_layers):
+      self.self_attn_layers.append(
+        MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout, proximal_bias=proximal_bias,
+                           proximal_init=proximal_init))
+      self.norm_layers_0.append(LayerNorm(hidden_channels))
+      self.ffn_layers.append(
+        FFN(hidden_channels, hidden_channels, filter_channels, kernel_size, p_dropout=p_dropout, causal=True))
+      self.norm_layers_1.append(LayerNorm(hidden_channels))
+
+  def forward(self, x, x_mask):
+    """
+    x: decoder input
+    h: encoder output
+    """
+    self_attn_mask = commons.subsequent_mask(x_mask.shape[2]).astype(dtype=x.dtype)
+    x = x * x_mask
+    for i in range(self.n_layers):
+      y = self.self_attn_layers[i](x, x, self_attn_mask)
+      y = self.drop(y)
+      x = self.norm_layers_0[i](x + y)
+
+      y = self.ffn_layers[i](x, x_mask)
+      y = self.drop(y)
+      x = self.norm_layers_1[i](x + y)
+    x = x * x_mask
+    return x
+
+
+class Encoder(nn.Layer):
+  def __init__(self, hidden_channels, filter_channels, n_heads, n_layers, kernel_size=1, p_dropout=0., window_size=4, **kwargs):
+    super().__init__()
+    self.hidden_channels = hidden_channels
+    self.filter_channels = filter_channels
+    self.n_heads = n_heads
+    self.n_layers = n_layers
+    self.kernel_size = kernel_size
+    self.p_dropout = p_dropout
+    self.window_size = window_size
+
+    self.drop = nn.Dropout(p_dropout)
+    self.attn_layers = nn.LayerList()
+    self.norm_layers_1 = nn.LayerList()
+    self.ffn_layers = nn.LayerList()
+    self.norm_layers_2 = nn.LayerList()
+    for i in range(self.n_layers):
+      self.attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout, window_size=window_size))
+      self.norm_layers_1.append(LayerNorm(hidden_channels))
+      self.ffn_layers.append(FFN(hidden_channels, hidden_channels, filter_channels, kernel_size, p_dropout=p_dropout))
+      self.norm_layers_2.append(LayerNorm(hidden_channels))
+
+  def forward(self, x, x_mask):
+    attn_mask = x_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
+    x = x * x_mask
+    for i in range(self.n_layers):
+      y = self.attn_layers[i](x, x, attn_mask)
+      y = self.drop(y)
+      x = self.norm_layers_1[i](x + y)
+      y = self.ffn_layers[i](x, x_mask)
+      y = self.drop(y)
+      x = self.norm_layers_2[i](x + y)
+    x = x * x_mask
+    return x
+
+
+class Decoder(nn.Layer):
+  def __init__(self, hidden_channels, filter_channels, n_heads, n_layers, kernel_size=1, p_dropout=0., proximal_bias=False, proximal_init=True, **kwargs):
+    super().__init__()
+    self.hidden_channels = hidden_channels
+    self.filter_channels = filter_channels
+    self.n_heads = n_heads
+    self.n_layers = n_layers
+    self.kernel_size = kernel_size
+    self.p_dropout = p_dropout
+    self.proximal_bias = proximal_bias
+    self.proximal_init = proximal_init
+
+    self.drop = nn.Dropout(p_dropout)
+    self.self_attn_layers = nn.LayerList()
+    self.norm_layers_0 = nn.LayerList()
+    self.encdec_attn_layers = nn.LayerList()
+    self.norm_layers_1 = nn.LayerList()
+    self.ffn_layers = nn.LayerList()
+    self.norm_layers_2 = nn.LayerList()
+    for i in range(self.n_layers):
+      self.self_attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout, proximal_bias=proximal_bias, proximal_init=proximal_init))
+      self.norm_layers_0.append(LayerNorm(hidden_channels))
+      self.encdec_attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout))
+      self.norm_layers_1.append(LayerNorm(hidden_channels))
+      self.ffn_layers.append(FFN(hidden_channels, hidden_channels, filter_channels, kernel_size, p_dropout=p_dropout, causal=True))
+      self.norm_layers_2.append(LayerNorm(hidden_channels))
+
+  def forward(self, x, x_mask, h, h_mask):
+    """
+    x: decoder input
+    h: encoder output
+    """
+    self_attn_mask = commons.subsequent_mask(x_mask.size(2)).astype(dtype=x.dtype)
+    encdec_attn_mask = h_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
+    x = x * x_mask
+    for i in range(self.n_layers):
+      y = self.self_attn_layers[i](x, x, self_attn_mask)
+      y = self.drop(y)
+      x = self.norm_layers_0[i](x + y)
+
+      y = self.encdec_attn_layers[i](x, h, encdec_attn_mask)
+      y = self.drop(y)
+      x = self.norm_layers_1[i](x + y)
+      
+      y = self.ffn_layers[i](x, x_mask)
+      y = self.drop(y)
+      x = self.norm_layers_2[i](x + y)
+    x = x * x_mask
+    return x
+
+
+class MultiHeadAttention(nn.Layer):
+  def __init__(self, channels, out_channels, n_heads, p_dropout=0., window_size=None, heads_share=True, block_length=None, proximal_bias=False, proximal_init=False):
+    super().__init__()
+    assert channels % n_heads == 0
+
+    self.channels = channels
+    self.out_channels = out_channels
+    self.n_heads = n_heads
+    self.p_dropout = p_dropout
+    self.window_size = window_size
+    self.heads_share = heads_share
+    self.block_length = block_length
+    self.proximal_bias = proximal_bias
+    self.proximal_init = proximal_init
+    self.attn = None
+
+    self.k_channels = channels // n_heads
+
+    self.conv_q = nn.Conv1D(channels, channels, 1,)# weight_attr=attr)
+    self.conv_k = nn.Conv1D(channels, channels, 1,)# weight_attr=attr)
+    self.conv_v = nn.Conv1D(channels, channels, 1,)# weight_attr=attr)
+    self.conv_o = nn.Conv1D(channels, out_channels, 1)
+    self.drop = nn.Dropout(p_dropout)
+
+    if window_size is not None:
+      n_heads_rel = 1 if heads_share else n_heads
+      rel_stddev = self.k_channels**-0.5
+
+      rand = paddle.randn((n_heads_rel, window_size * 2 + 1, self.k_channels)) * rel_stddev
+
+      self.emb_rel_k = paddle.create_parameter(rand.shape,'float32',None)
+      self.emb_rel_v = paddle.create_parameter(rand.shape,'float32',None)
+
+    #nn.init.xavier_uniform_(self.conv_q.weight)
+    #nn.init.xavier_uniform_(self.conv_k.weight)
+    #nn.init.xavier_uniform_(self.conv_v.weight)
+    if proximal_init:
+      with paddle.no_grad():
+        self.conv_k.weight = (self.conv_q.weight)
+        self.conv_k.bias = (self.conv_q.bias)
+      
+  def forward(self, x, c, attn_mask=None):
+    #print(x)
+    #print(self.conv_q.weight)
+    q = self.conv_q(x)
+    k = self.conv_k(c)
+    v = self.conv_v(c)
+
+    x, self.attn = self.attention(q, k, v, mask=attn_mask)
+
+    x = self.conv_o(x)
+    return x
+
+  @staticmethod
+  def _masked_fill(x, mask, value:float):
+    y = paddle.full(x.shape, value, x.dtype)
+    return paddle.where(mask, y, x)
+
+  def attention(self, query, key, value, mask=None):
+    # reshape [b, d, t] -> [b, n_h, t, d_k]
+    b, d, t_s, t_t = (*key.shape, query.shape[2])
+    query = query.reshape((b, self.n_heads, self.k_channels, t_t)).transpose([0,1,3,2])
+    key = key.reshape((b, self.n_heads, self.k_channels, t_s)).transpose([0,1,3,2])
+    value = value.reshape((b, self.n_heads, self.k_channels, t_s)).transpose([0,1,3,2])
+
+    scores = paddle.matmul(query / math.sqrt(self.k_channels), key.transpose([0,1,3,2])) # 0 1 2 3     -4 -3 -2 -1
+    if self.window_size is not None:
+      assert t_s == t_t, "Relative attention is only available for self-attention."
+
+      key_relative_embeddings = self._get_relative_embeddings(self.emb_rel_k, t_s)
+      rel_logits = self._matmul_with_relative_keys(query /math.sqrt(self.k_channels), key_relative_embeddings)
+      scores_local = self._relative_position_to_absolute_position(rel_logits)
+      scores = scores + scores_local
+    if self.proximal_bias:
+      assert t_s == t_t, "Proximal bias is only available for self-attention."
+      scores = scores + self._attention_bias_proximal(t_s).astype(dtype=scores.dtype)
+    if mask is not None:
+      scores = self._masked_fill(scores, mask == 0, -1e4)
+      if self.block_length is not None:
+        assert t_s == t_t, "Local attention is only available for self-attention."
+        block_mask = paddle.tril(paddle.triu(paddle.ones_like(scores), -self.block_length),self.block_length)
+        scores = self._masked_fill(scores, block_mask == 0, -1e4)
+    p_attn = F.softmax(scores, axis=-1) # [b, n_h, t_t, t_s]
+    p_attn = self.drop(p_attn)
+    output = paddle.matmul(p_attn, value)
+    if self.window_size is not None:
+      relative_weights = self._absolute_position_to_relative_position(p_attn)
+      value_relative_embeddings = self._get_relative_embeddings(self.emb_rel_v, t_s)
+      output = output + self._matmul_with_relative_values(relative_weights, value_relative_embeddings)
+    output = output.transpose([0,1,3,2]).reshape((b, d, t_t)) # [b, n_h, t_t, d_k] -> [b, d, t_t]
+    return output, p_attn
+
+  def _matmul_with_relative_values(self, x, y):
+    """
+    x: [b, h, l, m]
+    y: [h or 1, m, d]
+    ret: [b, h, l, d]
+    """
+    ret = paddle.matmul(x, y.unsqueeze(0))
+    return ret
+
+  def _matmul_with_relative_keys(self, x, y):
+    """
+    x: [b, h, l, d]
+    y: [h or 1, m, d]
+    ret: [b, h, l, m]
+    """
+    ret = paddle.matmul(x, y.unsqueeze(0).transpose([0,1,3,2]))
+    return ret
+
+  def _get_relative_embeddings(self, relative_embeddings, length):
+    max_relative_position = 2 * self.window_size + 1
+    # Pad first before slice to avoid using cond ops.
+    pad_length = max(length - (self.window_size + 1), 0)
+    slice_start_position = max((self.window_size + 1) - length, 0)
+    slice_end_position = slice_start_position + 2 * length - 1
+    if pad_length > 0:
+      padding = commons.convert_pad_shape([[0, 0], [pad_length, pad_length], [0, 0]])
+
+      padded_relative_embeddings = F.pad(
+          x = relative_embeddings.unsqueeze(0),
+          pad = padding[0:4]).squeeze(0)
+
+    else:
+      padded_relative_embeddings = relative_embeddings
+    used_relative_embeddings = padded_relative_embeddings[:,slice_start_position:slice_end_position]
+    return used_relative_embeddings
+
+  def _relative_position_to_absolute_position(self, x):
+    """
+    x: [b, h, l, 2*l-1]
+    ret: [b, h, l, l]
+    """
+    batch, heads, length, _ = x.shape
+    # Concat columns of pad to shift from relative to absolute indexing.
+    pad_shape = commons.convert_pad_shape([[0,0],[0,0],[0,0],[0,1]])
+    pad_shape = commons.fix_pad_shape(pad_shape, x)
+    x = F.pad(x, pad_shape)
+    # Concat extra elements so to add up to shape (len+1, 2*len-1).
+    x_flat = x.reshape([batch, heads, length * 2 * length])
+    pad_shape = commons.convert_pad_shape([[0,0],[0,0],[0,length-1]])
+    pad_shape = commons.fix_pad_shape(pad_shape,x_flat)
+    x_flat = F.pad(x_flat, pad_shape, data_format='NCL')
+    # Reshape and slice out the padded elements.
+    x_final = x_flat.reshape([batch, heads, length+1, 2*length-1])[:, :, :length, length-1:]
+    return x_final
+
+  def _absolute_position_to_relative_position(self, x):
+    """
+    x: [b, h, l, l]
+    ret: [b, h, l, 2*l-1]
+    """
+    batch, heads, length, _ = x.shape
+    # padd along column
+    pad_shape = commons.convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, length-1]])
+    pad_shape = commons.fix_pad_shape(pad_shape, x)
+    x = F.pad(x, pad_shape)
+    x_flat = x.reshape([batch, heads, length**2 + length*(length -1)])
+    # add 0's in the beginning that will skew the elements after reshape
+    pad_shape = commons.convert_pad_shape([[0, 0], [0, 0], [length, 0]])
+    pad_shape = commons.fix_pad_shape(pad_shape, x_flat)
+    x_flat = F.pad(x_flat, pad_shape, data_format='NCL')
+    x_final = x_flat.reshape([batch, heads, length, 2*length])[:,:,:,1:]
+    return x_final
+
+  def _attention_bias_proximal(self, length):
+    """Bias for self-attention to encourage attention to close positions.
+    Args:
+      length: an integer scalar.
+    Returns:
+      a Tensor with shape [1, 1, length, length]
+    """
+    r = paddle.arange(length, dtype=np.float32)
+    diff = paddle.unsqueeze(r, 0) - paddle.unsqueeze(r, 1)
+    return paddle.unsqueeze(paddle.unsqueeze(-paddle.log1p(paddle.abs(diff)), 0), 0)
+
+
+class FFN(nn.Layer):
+  def __init__(self, in_channels, out_channels, filter_channels, kernel_size, p_dropout=0., activation=None, causal=False):
+    super().__init__()
+    self.in_channels = in_channels
+    self.out_channels = out_channels
+    self.filter_channels = filter_channels
+    self.kernel_size = kernel_size
+    self.p_dropout = p_dropout
+    self.activation = activation
+    self.causal = causal
+
+    if causal:
+      self.padding = self._causal_padding
+    else:
+      self.padding = self._same_padding
+
+    self.conv_1 = nn.Conv1D(in_channels, filter_channels, kernel_size)
+    self.conv_2 = nn.Conv1D(filter_channels, out_channels, kernel_size)
+    self.drop = nn.Dropout(p_dropout)
+
+  def forward(self, x, x_mask):
+    x = x * x_mask
+    x = self.padding(x)
+    x = self.conv_1(x)
+    if self.activation == "gelu":
+      x = x * F.sigmoid(1.702 * x)
+    else:
+      x = F.relu(x)
+    x = self.drop(x)
+    x = x * x_mask
+    x = self.padding(x)
+    x = self.conv_2(x)
+    return x * x_mask
+  
+  def _causal_padding(self, x):
+    if self.kernel_size == 1:
+      return x
+    pad_l = self.kernel_size - 1
+    pad_r = 0
+    padding = [[0, 0], [0, 0], [pad_l, pad_r]]
+    pad_shape:list = commons.convert_pad_shape(padding)
+    pad_shape = commons.fix_pad_shape(pad_shape, x)
+    x = F.pad(x, pad_shape,data_format='NCL')
+    return x
+
+  def _same_padding(self, x):
+    if self.kernel_size == 1:
+      return x
+    pad_l = (self.kernel_size - 1) // 2
+    pad_r = self.kernel_size // 2
+    padding = [[0, 0], [0, 0], [pad_l, pad_r]]
+    pad_shape = commons.convert_pad_shape(padding)
+    pad_shape = commons.fix_pad_shape(pad_shape, x)
+    x = F.pad(x, pad_shape, data_format='NCL')
+    return x

modules/commons.py

@@ -0,0 +1,192 @@
+import math
+import numpy as np
+import paddle
+from paddle import nn
+from paddle.nn import functional as F
+
+def slice_pitch_segments(x, ids_str, segment_size=4):
+  ret = paddle.zeros_like(x[:, :segment_size])
+  for i in range(x.shape[0]):
+    idx_str = ids_str[i]
+    idx_end = idx_str + segment_size
+    ret[i] = x[i, idx_str:idx_end]
+  return ret
+
+def rand_slice_segments_with_pitch(x, pitch, x_lengths=None, segment_size=4):
+  b, d, t = x.shape
+  if x_lengths is None:
+    x_lengths = t
+  ids_str_max = x_lengths - segment_size + 1
+  ids_str = (paddle.rand([b]) * ids_str_max.astype('float32')).astype(dtype='int64')
+  ret = slice_segments(x, ids_str, segment_size)
+  ret_pitch = slice_pitch_segments(pitch, ids_str, segment_size)
+  return ret, ret_pitch, ids_str
+
+def init_weights(m, mean=0.0, std=0.01):
+    classname = m.__class__.__name__
+    if classname.find("Conv") != -1:
+      m.weight.data.normal_(mean, std)
+
+
+def get_padding(kernel_size, dilation=1):
+  return int((kernel_size*dilation - dilation)/2)
+
+
+def convert_pad_shape(pad_shape):
+  l = pad_shape[::-1]
+  pad_shape = paddle.to_tensor([item for sublist in l for item in sublist],).flatten().astype('int32')
+  return pad_shape
+
+
+def intersperse(lst, item):
+  result = [item] * (len(lst) * 2 + 1)
+  result[1::2] = lst
+  return result
+
+
+def kl_divergence(m_p, logs_p, m_q, logs_q):
+  """KL(P||Q)"""
+  kl = (logs_q - logs_p) - 0.5
+  kl += 0.5 * (paddle.exp(2. * logs_p) + ((m_p - m_q)**2)) * paddle.exp(-2. * logs_q)
+  return kl
+
+
+def rand_gumbel(shape):
+  """Sample from the Gumbel distribution, protect from overflows."""
+  uniform_samples = paddle.rand(shape) * 0.99998 + 0.00001
+  return -paddle.log(-paddle.log(uniform_samples))
+
+
+def rand_gumbel_like(x):
+  g = rand_gumbel(x.shape).astype(dtype=x.dtype)
+  return g
+
+
+def slice_segments(x, ids_str, segment_size=4):
+  ret = paddle.zeros_like(x[:, :, :segment_size])
+  for i in range(x.shape[0]):
+    idx_str = ids_str[i]
+    idx_end = idx_str + segment_size
+    ret[i] = x[i, :, idx_str:idx_end]
+  return ret
+
+
+def rand_slice_segments(x, x_lengths=None, segment_size=4):
+  b, d, t = x.size()
+  if x_lengths is None:
+    x_lengths = t
+  ids_str_max = x_lengths - segment_size + 1
+  ids_str = (paddle.rand([b]) * ids_str_max).astype('int64')
+  ret = slice_segments(x, ids_str, segment_size)
+  return ret, ids_str
+
+
+def rand_spec_segments(x, x_lengths=None, segment_size=4):
+  b, d, t = x.size()
+  if x_lengths is None:
+    x_lengths = t
+  ids_str_max = x_lengths - segment_size
+  ids_str = (paddle.rand([b]) * ids_str_max).astype('int64')
+  ret = slice_segments(x, ids_str, segment_size)
+  return ret, ids_str
+
+
+def get_timing_signal_1d(
+    length, channels, min_timescale=1.0, max_timescale=1.0e4):
+  position = paddle.arange(length, dtype=np.float32)
+  num_timescales = channels // 2
+  log_timescale_increment = (
+      math.log(float(max_timescale) / float(min_timescale)) /
+      (num_timescales - 1))
+  inv_timescales = min_timescale * paddle.exp(
+      paddle.arange(num_timescales, dtype=np.float32) * -log_timescale_increment)
+  scaled_time = position.unsqueeze(0) * inv_timescales.unsqueeze(1)
+  signal = paddle.concat([paddle.sin(scaled_time), paddle.cos(scaled_time)], 0)
+  signal = F.pad(signal, [0, 0, 0, channels % 2])
+  signal = signal.reshape((1, channels, length))
+  return signal
+
+
+def add_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4):
+  b, channels, length = x.shape
+  signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)
+  return x + signal.astype(dtype=x.dtype)
+
+
+def cat_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4, axis=1):
+  b, channels, length = x.size()
+  signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)
+  return paddle.concat([x, signal.astype(dtype=x.dtype)], axis)
+
+
+def subsequent_mask(length):
+  mask = paddle.tril(paddle.ones((length, length))).unsqueeze(0).unsqueeze(0)
+  return mask
+
+
+#@paddle.jit.to_static # @torch.jit.script
+def fused_add_tanh_sigmoid_multiply(input_a, input_b, n_channels):
+  n_channels_int = n_channels[0]
+  in_act = input_a + input_b
+  t_act = paddle.tanh(in_act[:, :n_channels_int, :])
+  s_act = paddle.nn.functional.sigmoid(in_act[:, n_channels_int:, :])
+  print(t_act)
+  print(s_act)
+  acts = t_act * s_act
+  return acts
+
+def fix_pad_shape(pad_shape:paddle.Tensor, pad_tensor) -> paddle.Tensor: # 飞桨里面的padding函数对pad_shape有比较严格的要求，需要自己修正一下~~~
+  if len(pad_tensor.shape) == 3:
+    return pad_shape[0:2].astype('int32')
+  elif len(pad_tensor.shape) == 4:
+    return pad_shape[0:4].astype('int32')
+  elif len(pad_tensor.shape) == 5:
+    return pad_shape[0:6].astype('int32')
+  return pad_shape.astype('int32')
+
+def shift_1d(x):
+  x = F.pad(x, convert_pad_shape([[0, 0], [0, 0], [1, 0]]))[:, :, :-1]
+  return x
+
+
+def sequence_mask(length:paddle.Tensor, max_length=None):
+  if max_length is None:
+    max_length = length.max()
+  x = paddle.arange(max_length, dtype=length.dtype)
+  return x.unsqueeze(0) < length.unsqueeze(1)
+
+
+def generate_path(duration, mask):
+  """
+  duration: [b, 1, t_x]
+  mask: [b, 1, t_y, t_x]
+  """
+  device = duration.device
+  
+  b, _, t_y, t_x = mask.shape
+  cum_duration = paddle.cumsum(duration, -1)
+  
+  cum_duration_flat = cum_duration.reshape((b * t_x))
+  path = sequence_mask(cum_duration_flat, t_y).astype(mask.dtype)
+  path = path.reshape((b, t_x, t_y))
+  path = path - F.pad(path, convert_pad_shape([[0, 0], [1, 0], [0, 0]]))[:, :-1]
+  path = path.unsqueeze(1).transpose([0,1,3,2]) * mask
+  return path
+
+
+def clip_grad_value_(parameters, clip_value, norm_type=2):
+  if isinstance(parameters, paddle.Tensor):
+    parameters = [parameters]
+  parameters = list(filter(lambda p: p.grad is not None, parameters))
+  norm_type = float(norm_type)
+  if clip_value is not None:
+    clip_value = float(clip_value)
+
+  total_norm = 0
+  for p in parameters:
+    param_norm = paddle.to_tensor(p.grad).norm(norm_type)
+    total_norm += param_norm.item() ** norm_type
+    if clip_value is not None:
+      paddle.to_tensor(p.grad).clip_(min=-clip_value, max=clip_value)
+  total_norm = total_norm ** (1. / norm_type)
+  return total_norm

modules/losses.py

@@ -0,0 +1,61 @@
+import paddle 
+from paddle.nn import functional as F
+
+import modules.commons as commons
+
+
+def feature_loss(fmap_r, fmap_g):
+  loss = 0
+  for dr, dg in zip(fmap_r, fmap_g):
+    for rl, gl in zip(dr, dg):
+      rl = rl.astype('float32').detach()
+      gl = gl.astype('float32')
+      loss += paddle.mean(paddle.abs(rl - gl))
+
+  return loss * 2 
+
+
+def discriminator_loss(disc_real_outputs, disc_generated_outputs):
+  loss = 0
+  r_losses = []
+  g_losses = []
+  for dr, dg in zip(disc_real_outputs, disc_generated_outputs):
+    dr = dr.astype('float32')
+    dg = dg.astype('float32')
+    r_loss = paddle.mean((1-dr)**2)
+    g_loss = paddle.mean(dg**2)
+    loss += (r_loss + g_loss)
+    r_losses.append(r_loss.item())
+    g_losses.append(g_loss.item())
+
+  return loss, r_losses, g_losses
+
+
+def generator_loss(disc_outputs):
+  loss = 0
+  gen_losses = []
+  for dg in disc_outputs:
+    dg = dg.astype('float32')
+    l = paddle.mean((1-dg)**2)
+    gen_losses.append(l)
+    loss += l
+
+  return loss, gen_losses
+
+
+def kl_loss(z_p, logs_q, m_p, logs_p, z_mask):
+  """
+  z_p, logs_q: [b, h, t_t]
+  m_p, logs_p: [b, h, t_t]
+  """
+  z_p = z_p.astype('float32')
+  logs_q = logs_q.astype('float32')
+  m_p = m_p.astype('float32')
+  logs_p = logs_p.astype('float32')
+  z_mask = z_mask.astype('float32')
+  #print(logs_p)
+  kl = logs_p - logs_q - 0.5
+  kl += 0.5 * ((z_p - m_p)**2) * paddle.exp(-2. * logs_p)
+  kl = paddle.sum(kl * z_mask)
+  l = kl / paddle.sum(z_mask)
+  return l

modules/mel_processing.py

@@ -0,0 +1,111 @@
+import math
+import os
+import random
+import paddle
+from paddle import nn
+import paddle.nn.functional as F
+import numpy as np
+import librosa
+import librosa.util as librosa_util
+from librosa.util import normalize, pad_center, tiny
+from scipy.signal import get_window
+from scipy.io.wavfile import read
+from librosa.filters import mel as librosa_mel_fn
+
+MAX_WAV_VALUE = 32768.0
+
+
+def dynamic_range_compression_torch(x, C=1, clip_val=1e-5):
+    """
+    PARAMS
+    ------
+    C: compression factor
+    """
+    return paddle.log(paddle.clip(x, min=clip_val) * C)
+
+
+def dynamic_range_decompression_torch(x, C=1):
+    """
+    PARAMS
+    ------
+    C: compression factor used to compress
+    """
+    return paddle.exp(x) / C
+
+
+def spectral_normalize_torch(magnitudes):
+    output = dynamic_range_compression_torch(magnitudes)
+    return output
+
+
+def spectral_de_normalize_torch(magnitudes):
+    output = dynamic_range_decompression_torch(magnitudes)
+    return output
+
+
+mel_basis = {}
+hann_window = {}
+
+
+def spectrogram_torch(y, n_fft, sampling_rate, hop_size, win_size, center=False):
+    if paddle.min(y) < -1.:
+        print('min value is ', paddle.min(y))
+    if paddle.max(y) > 1.:
+        print('max value is ', paddle.max(y))
+
+    global hann_window
+    dtype_device = str(y.dtype) + '_' + str(str(y.place)[6:-1])
+    wnsize_dtype_device = str(win_size) + '_' + dtype_device
+    if wnsize_dtype_device not in hann_window:
+        hann_window[wnsize_dtype_device] = paddle.audio.functional.get_window('hann',win_size).astype(y.dtype)
+
+    y = paddle.nn.functional.pad(y.unsqueeze(1), (int((n_fft-hop_size)/2), int((n_fft-hop_size)/2)), mode='reflect', data_format='NCL')
+    y = y.squeeze(1)
+
+    spec = paddle.signal.stft(y, n_fft, hop_length=hop_size, win_length=win_size, window=hann_window[wnsize_dtype_device],
+                      center=center, pad_mode='reflect', normalized=False, onesided=True)
+    spec = paddle.as_real(spec)
+    spec = paddle.sqrt(spec.pow(2).sum(-1) + 1e-6)
+    return spec
+
+
+def spec_to_mel_torch(spec, n_fft, num_mels, sampling_rate, fmin, fmax):
+    global mel_basis
+    dtype_device = str(spec.dtype) + '_' + str(spec.place)[6:-1]
+    fmax_dtype_device = str(fmax) + '_' + dtype_device
+    if fmax_dtype_device not in mel_basis:
+        mel = librosa_mel_fn(sr=sampling_rate, n_fft=n_fft, n_mels=num_mels, fmin=fmin, fmax=fmax)
+        mel_basis[fmax_dtype_device] = paddle.to_tensor(mel).astype(spec.dtype)
+    spec = paddle.matmul(mel_basis[fmax_dtype_device], spec)
+    spec = spectral_normalize_torch(spec)
+    return spec
+
+
+def mel_spectrogram_torch(y, n_fft, num_mels, sampling_rate, hop_size, win_size, fmin, fmax, center=False):
+    if paddle.min(y) < -1.:
+        print('min value is ', paddle.min(y))
+    if paddle.max(y) > 1.:
+        print('max value is ', paddle.max(y))
+
+    global mel_basis, hann_window
+    dtype_device = str(y.dtype) + '_' + str(y.place)[6:-1]
+    fmax_dtype_device = str(fmax) + '_' + dtype_device
+    wnsize_dtype_device = str(win_size) + '_' + dtype_device
+    if fmax_dtype_device not in mel_basis:
+        mel = librosa_mel_fn(sr=sampling_rate, n_fft=n_fft, n_mels=num_mels, fmin=fmin, fmax=fmax)
+        mel_basis[fmax_dtype_device] = paddle.to_tensor(mel).astype(y.dtype)
+    if wnsize_dtype_device not in hann_window:
+        hann_window[wnsize_dtype_device] = paddle.audio.functional.get_window('hann',win_size).astype(y.dtype)
+
+    y = paddle.nn.functional.pad(y.unsqueeze(1), (int((n_fft-hop_size)/2), int((n_fft-hop_size)/2)), mode='reflect',data_format = 'NCL')
+    y = y.squeeze(1)
+
+    spec = paddle.signal.stft(y, n_fft, hop_length=hop_size, win_length=win_size, window=hann_window[wnsize_dtype_device],
+                      center=center, pad_mode='reflect', normalized=False, onesided=True)
+    spec = paddle.as_real(spec)
+    spec = paddle.sqrt(spec.pow(2).sum(-1) + 1e-6)
+
+    spec = paddle.matmul(mel_basis[fmax_dtype_device], spec)
+    spec = spectral_normalize_torch(spec)
+
+    return spec

modules/modules.py

@@ -0,0 +1,351 @@
+import copy
+import math
+import numpy as np
+import scipy
+import paddle
+from paddle import nn
+from paddle.nn import functional as F
+
+from paddle.nn import Conv1D, Conv1DTranspose, AvgPool1D, Conv2D
+from paddle.nn.utils import weight_norm, remove_weight_norm
+
+import modules.commons as commons
+from modules.commons import init_weights, get_padding
+
+
+LRELU_SLOPE = 0.1
+
+
+class LayerNorm(nn.Layer):
+  def __init__(self, channels, eps=1e-5):
+    super().__init__()
+    self.channels = channels
+    self.eps = eps
+
+    self.gamma = paddle.create_parameter([channels],'float32','modules_Layer_Norm_gamma',\
+        paddle.ParamAttr(initializer = paddle.nn.initializer.Constant(value=1.0))) # ones,shape = [channels]
+    self.beta = paddle.create_parameter([channels],'float32','modules_Layer_Norm_beta',\
+        paddle.ParamAttr(initializer = paddle.nn.initializer.Constant(value=0.0))) # zeros,shape = [channels]
+
+  def forward(self, x):
+    x = x.transpose([0,2,1])#x.transpose(1, -1)
+    x = F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps)
+    return x.transpose([0,2,1])#x.transpose(1, -1)
+
+ 
+class ConvReluNorm(nn.Layer):
+  def __init__(self, in_channels, hidden_channels, out_channels, kernel_size, n_layers, p_dropout):
+    super().__init__()
+    self.in_channels = in_channels
+    self.hidden_channels = hidden_channels
+    self.out_channels = out_channels
+    self.kernel_size = kernel_size
+    self.n_layers = n_layers
+    self.p_dropout = p_dropout
+    assert n_layers > 1, "Number of layers should be larger than 0."
+
+    self.conv_layers = nn.LayerList()
+    self.norm_layers = nn.LayerList()
+    self.conv_layers.append(nn.Conv1D(in_channels, hidden_channels, kernel_size, padding=kernel_size//2))
+    self.norm_layers.append(LayerNorm(hidden_channels))
+    self.relu_drop = nn.Sequential(
+        nn.ReLU(),
+        nn.Dropout(p_dropout))
+    for _ in range(n_layers-1):
+      self.conv_layers.append(nn.Conv1D(hidden_channels, hidden_channels, kernel_size, padding=kernel_size//2))
+      self.norm_layers.append(LayerNorm(hidden_channels))
+    att = paddle.ParamAttr('modules_ConvReluNorm_att',initializer = paddle.nn.initializer.Constant(value=0.0)) # น้มใ
+    self.proj = nn.Conv1D(hidden_channels, out_channels, 1, weight_attr=att, bias_attr=att)
+    #self.proj.weight.data.zero_()
+    #self.proj.bias.data.zero_()
+
+  def forward(self, x, x_mask):
+    x_org = x
+    for i in range(self.n_layers):
+      x = self.conv_layers[i](x * x_mask)
+      x = self.norm_layers[i](x)
+      x = self.relu_drop(x)
+    x = x_org + self.proj(x)
+    return x * x_mask
+
+
+class DDSConv(nn.Layer):
+  """
+  Dialted and Depth-Separable Convolution
+  """
+  def __init__(self, channels, kernel_size, n_layers, p_dropout=0.):
+    super().__init__()
+    self.channels = channels
+    self.kernel_size = kernel_size
+    self.n_layers = n_layers
+    self.p_dropout = p_dropout
+
+    self.drop = nn.Dropout(p_dropout)
+    self.convs_sep = nn.LayerList()
+    self.convs_1x1 = nn.LayerList()
+    self.norms_1 = nn.LayerList()
+    self.norms_2 = nn.LayerList()
+    for i in range(n_layers):
+      dilation = kernel_size ** i
+      padding = (kernel_size * dilation - dilation) // 2
+      self.convs_sep.append(nn.Conv1D(channels, channels, kernel_size, 
+          groups=channels, dilation=dilation, padding=padding
+      ))
+      self.convs_1x1.append(nn.Conv1D(channels, channels, 1))
+      self.norms_1.append(LayerNorm(channels))
+      self.norms_2.append(LayerNorm(channels))
+
+  def forward(self, x, x_mask, g=None):
+    if g is not None:
+      x = x + g
+    for i in range(self.n_layers):
+      y = self.convs_sep[i](x * x_mask)
+      y = self.norms_1[i](y)
+      y = F.gelu(y)
+      y = self.convs_1x1[i](y)
+      y = self.norms_2[i](y)
+      y = F.gelu(y)
+      y = self.drop(y)
+      x = x + y
+    return x * x_mask
+
+
+class WN(paddle.nn.Layer):
+  def __init__(self, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=0, p_dropout=0):
+    super(WN, self).__init__()
+    assert(kernel_size % 2 == 1)
+    self.hidden_channels =hidden_channels
+    self.kernel_size = kernel_size,
+    self.dilation_rate = dilation_rate
+    self.n_layers = n_layers
+    self.gin_channels = gin_channels
+    self.p_dropout = p_dropout
+
+    self.in_layers = paddle.nn.LayerList()
+    self.res_skip_layers = paddle.nn.LayerList()
+    self.drop = nn.Dropout(p_dropout)
+
+    if gin_channels != 0:
+      cond_layer = paddle.nn.Conv1D(gin_channels, 2*hidden_channels*n_layers, 1)
+      self.cond_layer = paddle.nn.utils.weight_norm(cond_layer, name='weight')
+
+    for i in range(n_layers):
+      dilation = dilation_rate ** i
+      padding = int((kernel_size * dilation - dilation) / 2)
+      in_layer = paddle.nn.Conv1D(hidden_channels, 2*hidden_channels, kernel_size,
+                                 dilation=dilation, padding=padding)
+      in_layer = paddle.nn.utils.weight_norm(in_layer, name='weight')
+      self.in_layers.append(in_layer)
+
+      # last one is not necessary
+      if i < n_layers - 1:
+        res_skip_channels = 2 * hidden_channels
+      else:
+        res_skip_channels = hidden_channels
+
+      res_skip_layer = paddle.nn.Conv1D(hidden_channels, res_skip_channels, 1)
+      res_skip_layer = paddle.nn.utils.weight_norm(res_skip_layer, name='weight')
+      self.res_skip_layers.append(res_skip_layer)
+
+  def forward(self, x, x_mask, g=None, **kwargs):
+    output = paddle.zeros_like(x,name = 'module_WN_forward_output')
+
+    if g is not None:
+      g = self.cond_layer(g)
+
+    for i in range(self.n_layers):
+      x_in = self.in_layers[i](x)
+      if g is not None:
+        cond_offset = i * 2 * self.hidden_channels
+        g_l = g[:,cond_offset:cond_offset+2*self.hidden_channels,:]
+      else:
+        g_l = paddle.zeros_like(x_in,name = 'module_WN_forward_gl')
+
+      input_a=x_in; input_b=g_l
+      n_channels_int = self.hidden_channels
+      in_act = input_a + input_b
+      t_act = paddle.tanh(in_act[:, :n_channels_int, :])
+      s_act = paddle.nn.functional.sigmoid(in_act[:, n_channels_int:, :])
+      acts = t_act * s_act
+
+      acts = self.drop(acts)
+
+      res_skip_acts = self.res_skip_layers[i](acts)
+      if i < self.n_layers - 1:
+        res_acts = res_skip_acts[:,:self.hidden_channels,:]
+        x = (x + res_acts) * x_mask
+        output = output + res_skip_acts[:,self.hidden_channels:,:]
+      else:
+        output = output + res_skip_acts
+    return output * x_mask
+
+  def remove_weight_norm(self):
+    if self.gin_channels != 0:
+      paddle.nn.utils.remove_weight_norm(self.cond_layer)
+    for l in self.in_layers:
+      paddle.nn.utils.remove_weight_norm(l)
+    for l in self.res_skip_layers:
+     paddle.nn.utils.remove_weight_norm(l)
+
+
+class ResBlock1(paddle.nn.Layer):
+    def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5)):
+        super(ResBlock1, self).__init__()
+        self.convs1 = nn.LayerList([
+            weight_norm(Conv1D(channels, channels, kernel_size, 1, dilation=dilation[0],
+                               padding=get_padding(kernel_size, dilation[0]))),
+            weight_norm(Conv1D(channels, channels, kernel_size, 1, dilation=dilation[1],
+                               padding=get_padding(kernel_size, dilation[1]))),
+            weight_norm(Conv1D(channels, channels, kernel_size, 1, dilation=dilation[2],
+                               padding=get_padding(kernel_size, dilation[2])))
+        ])
+        self.convs1.apply(init_weights)
+
+        self.convs2 = nn.LayerList([
+            weight_norm(Conv1D(channels, channels, kernel_size, 1, dilation=1,
+                               padding=get_padding(kernel_size, 1))),
+            weight_norm(Conv1D(channels, channels, kernel_size, 1, dilation=1,
+                               padding=get_padding(kernel_size, 1))),
+            weight_norm(Conv1D(channels, channels, kernel_size, 1, dilation=1,
+                               padding=get_padding(kernel_size, 1)))
+        ])
+        self.convs2.apply(init_weights)
+
+    def forward(self, x, x_mask=None):
+        for c1, c2 in zip(self.convs1, self.convs2):
+            xt = F.leaky_relu(x, LRELU_SLOPE)
+            if x_mask is not None:
+                xt = xt * x_mask
+            xt = c1(xt)
+            xt = F.leaky_relu(xt, LRELU_SLOPE)
+            if x_mask is not None:
+                xt = xt * x_mask
+            xt = c2(xt)
+            x = xt + x
+        if x_mask is not None:
+            x = x * x_mask
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.convs1:
+            remove_weight_norm(l)
+        for l in self.convs2:
+            remove_weight_norm(l)
+
+
+class ResBlock2(paddle.nn.Layer):
+    def __init__(self, channels, kernel_size=3, dilation=(1, 3)):
+        super(ResBlock2, self).__init__()
+        self.convs = nn.LayerList([
+            weight_norm(Conv1D(channels, channels, kernel_size, 1, dilation=dilation[0],
+                               padding=get_padding(kernel_size, dilation[0]))),
+            weight_norm(Conv1D(channels, channels, kernel_size, 1, dilation=dilation[1],
+                               padding=get_padding(kernel_size, dilation[1])))
+        ])
+        self.convs.apply(init_weights)
+
+    def forward(self, x, x_mask=None):
+        for c in self.convs:
+            xt = F.leaky_relu(x, LRELU_SLOPE)
+            if x_mask is not None:
+                xt = xt * x_mask
+            xt = c(xt)
+            x = xt + x
+        if x_mask is not None:
+            x = x * x_mask
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.convs:
+            remove_weight_norm(l)
+
+
+class Log(nn.Layer):
+
+  def forward(self, x, x_mask, reverse=False, **kwargs):
+    if not reverse:
+      y = paddle.log(paddle.clip(x, 1e-5)) * x_mask
+      logdet = paddle.sum(-y, [1, 2])
+      return y, logdet
+    else:
+      x = paddle.exp(x) * x_mask
+      return x
+    
+
+class Flip(nn.Layer):
+  def forward(self, x, *args, reverse=False, **kwargs):
+    x = paddle.flip(x, [1])
+    if not reverse:
+      logdet = paddle.zeros([x.shape[0]]).astype(x.dtype)
+      return x, logdet
+    else:
+      return x
+
+
+class ElementwiseAffine(nn.Layer):
+  def __init__(self, channels):
+    super().__init__()
+    self.channels = channels
+    self.m = paddle.create_parameter([channels,1],'float32',None,\
+        paddle.ParamAttr(initializer = paddle.nn.initializer.Constant(value=0.0)))
+    self.logs = paddle.create_parameter([channels,1],'float32',None,\
+        paddle.ParamAttr(initializer = paddle.nn.initializer.Constant(value=0.0)))
+
+  def forward(self, x, x_mask, reverse=False, **kwargs):
+    if not reverse:
+      y = self.m + paddle.exp(self.logs) * x
+      y = y * x_mask
+      logdet = paddle.sum(self.logs * x_mask, [1,2])
+      return y, logdet
+    else:
+      x = (x - self.m) * paddle.exp(-self.logs) * x_mask
+      return x
+
+
+class ResidualCouplingLayer(nn.Layer):
+  def __init__(self,
+      channels,
+      hidden_channels,
+      kernel_size,
+      dilation_rate,
+      n_layers,
+      p_dropout=0,
+      gin_channels=0,
+      mean_only=False):
+    assert channels % 2 == 0, "channels should be divisible by 2"
+    super().__init__()
+    self.channels = channels
+    self.hidden_channels = hidden_channels
+    self.kernel_size = kernel_size
+    self.dilation_rate = dilation_rate
+    self.n_layers = n_layers
+    self.half_channels = channels // 2
+    self.mean_only = mean_only
+
+    self.pre = nn.Conv1D(self.half_channels, hidden_channels, 1)
+    self.enc = WN(hidden_channels, kernel_size, dilation_rate, n_layers, p_dropout=p_dropout, gin_channels=gin_channels)
+    att = paddle.ParamAttr(initializer = paddle.nn.initializer.Constant(value=0.0)) # น้มใ
+    self.post = nn.Conv1D(hidden_channels, self.half_channels * (2 - mean_only), 1,weight_attr=att, bias_attr=att)
+    #self.post.weight.data.zero_()
+    #self.post.bias.data.zero_()
+
+  def forward(self, x, x_mask, g=None, reverse=False):
+    x0, x1 = paddle.split(x, [self.half_channels]*2, 1)
+    h = self.pre(x0) * x_mask
+    h = self.enc(h, x_mask, g=g)
+    stats = self.post(h) * x_mask
+    if not self.mean_only:
+      m, logs = paddle.split(stats, [self.half_channels]*2, 1)
+    else:
+      m = stats
+      logs = paddle.zeros_like(m)
+
+    if not reverse:
+      x1 = m + x1 * paddle.exp(logs) * x_mask
+      x = paddle.concat([x0, x1], 1)
+      logdet = paddle.sum(logs, [1,2])
+      return x, logdet
+    else:
+      x1 = (x1 - m) * paddle.exp(-logs) * x_mask
+      x = paddle.concat([x0, x1], 1)
+      return x

output_2stems/blue-instrumental.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b9cdb5adc7ccd29f82f8f0d13adbdc83d4f8d9e56ea6f56d206f44e06e6ed690
+size 2704274

output_2stems/blue-vocals.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:59edf5461750769e48eb8d72ab41bfbe92a2483d6fcf725a87fa52659d8400ac
+size 2704274

output_2stems/temp-instrumental.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3583c87e8e45021d925579ef0f229ec88c6c73846509c6d08c03db0a18faba5e
+size 2704274

output_2stems/temp-vocals.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:80b9fe08d766a205dc0e60f2fe80b951ec95c35a2c32fbc2bb72e98a7b3ccfff
+size 2704274

paddle_infer_shape.py

@@ -0,0 +1,88 @@
+# 这个脚本文件用于更改静态图的静态Shape为动态Shape，从飞桨的GitHub上面复制的
+import argparse
+
+
+def process_old_ops_desc(program):
+    for i in range(len(program.blocks[0].ops)):
+        if program.blocks[0].ops[i].type == "matmul":
+            if not program.blocks[0].ops[i].has_attr("head_number"):
+                program.blocks[0].ops[i]._set_attr("head_number", 1)
+
+
+def infer_shape(program, input_shape_dict):
+    import paddle
+    paddle.enable_static()
+    import paddle.fluid as fluid
+
+    OP_WITHOUT_KERNEL_SET = {
+        'feed', 'fetch', 'recurrent', 'go', 'rnn_memory_helper_grad',
+        'conditional_block', 'while', 'send', 'recv', 'listen_and_serv',
+        'fl_listen_and_serv', 'ncclInit', 'select', 'checkpoint_notify',
+        'gen_bkcl_id', 'c_gen_bkcl_id', 'gen_nccl_id', 'c_gen_nccl_id',
+        'c_comm_init', 'c_sync_calc_stream', 'c_sync_comm_stream',
+        'queue_generator', 'dequeue', 'enqueue', 'heter_listen_and_serv',
+        'c_wait_comm', 'c_wait_compute', 'c_gen_hccl_id', 'c_comm_init_hccl',
+        'copy_cross_scope'
+    }
+    model_version = program.desc._version()
+    paddle_version = paddle.__version__
+    major_ver = model_version // 1000000
+    minor_ver = (model_version - major_ver * 1000000) // 1000
+    patch_ver = model_version - major_ver * 1000000 - minor_ver * 1000
+    model_version = "{}.{}.{}".format(major_ver, minor_ver, patch_ver)
+    if model_version != paddle_version:
+        print(
+            "[WARNING] The model is saved by paddlepaddle v{}, but now your paddlepaddle is version of {}, this difference may cause error, it is recommend you reinstall a same version of paddlepaddle for this model".
+            format(model_version, paddle_version))
+    for k, v in input_shape_dict.items():
+        program.blocks[0].var(k).desc.set_shape(v)
+    for i in range(len(program.blocks)):
+        for j in range(len(program.blocks[0].ops)):
+            if program.blocks[i].ops[j].type in OP_WITHOUT_KERNEL_SET:
+                continue
+            program.blocks[i].ops[j].desc.infer_shape(program.blocks[i].desc)
+
+
+def parse_arguments():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        '--model_dir',
+        required=True,
+        help='Path of directory saved the input model.')
+    parser.add_argument(
+        '--model_filename', required=True, help='The input model file name.')
+    parser.add_argument(
+        '--params_filename', required=True, help='The parameters file name.')
+    parser.add_argument(
+        '--save_dir',
+        required=True,
+        help='Path of directory to save the new exported model.')
+    parser.add_argument(
+        '--input_shape_dict', required=True, help="The new shape information.")
+    return parser.parse_args()
+
+
+if __name__ == '__main__':
+    args = parse_arguments()
+    import paddle
+    paddle.enable_static()
+    import paddle.fluid as fluid
+    input_shape_dict_str = args.input_shape_dict
+    input_shape_dict = eval(input_shape_dict_str)
+    print("Start to load paddle model...")
+    exe = fluid.Executor(fluid.CPUPlace())
+    [prog, ipts, outs] = fluid.io.load_inference_model(
+        args.model_dir,
+        exe,
+        model_filename=args.model_filename,
+        params_filename=args.params_filename)
+    process_old_ops_desc(prog)
+    infer_shape(prog, input_shape_dict)
+    fluid.io.save_inference_model(
+        args.save_dir,
+        ipts,
+        outs,
+        exe,
+        prog,
+        model_filename=args.model_filename,
+        params_filename=args.params_filename)

preprocess_flist_config.py

@@ -0,0 +1,84 @@
+import os
+import argparse
+import re
+
+from tqdm import tqdm
+from random import shuffle
+import json
+import wave
+
+config_template = json.load(open("configs/config.json"))
+
+pattern = re.compile(r'^[\.a-zA-Z0-9_\/]+$')
+
+def get_wav_duration(file_path):
+    with wave.open(file_path, 'rb') as wav_file:
+        # 获取音频帧数
+        n_frames = wav_file.getnframes()
+        # 获取采样率
+        framerate = wav_file.getframerate()
+        # 计算时长（秒）
+        duration = n_frames / float(framerate)
+    return duration
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--train_list", type=str, default="./filelists/train.txt", help="path to train list")
+    parser.add_argument("--val_list", type=str, default="./filelists/val.txt", help="path to val list")
+    parser.add_argument("--test_list", type=str, default="./filelists/test.txt", help="path to test list")
+    parser.add_argument("--source_dir", type=str, default="./dataset/44k", help="path to source dir")
+    args = parser.parse_args()
+    
+    train = []
+    val = []
+    test = []
+    idx = 0
+    spk_dict = {}
+    spk_id = 0
+    for speaker in tqdm(os.listdir(args.source_dir)):
+        spk_dict[speaker] = spk_id
+        spk_id += 1
+        wavs = ["/".join([args.source_dir, speaker, i]) for i in os.listdir(os.path.join(args.source_dir, speaker))]
+        new_wavs = []
+        for file in wavs:
+            if not file.endswith("wav"):
+                continue
+            if not pattern.match(file):
+                print(f"警告：文件名{file}中包含非字母数字下划线，可能会导致错误。（也可能不会）")
+                print('梅花自己测试发现是会出现问题的。')
+            if get_wav_duration(file) < 0.3:
+                print("跳过太短的音频：", file)
+                continue
+            new_wavs.append(file)
+        wavs = new_wavs
+        shuffle(wavs)
+        train += wavs[2:-2]
+        val += wavs[:2]
+        test += wavs[-2:]
+
+    shuffle(train)
+    shuffle(val)
+    shuffle(test)
+            
+    print("写入：", args.train_list)
+    with open(args.train_list, "w") as f:
+        for fname in tqdm(train):
+            wavpath = fname
+            f.write(wavpath + "\n")
+        
+    print("写入：", args.val_list)
+    with open(args.val_list, "w") as f:
+        for fname in tqdm(val):
+            wavpath = fname
+            f.write(wavpath + "\n")
+            
+    print("写入：", args.test_list)
+    with open(args.test_list, "w") as f:
+        for fname in tqdm(test):
+            wavpath = fname
+            f.write(wavpath + "\n")
+
+    config_template["spk"] = spk_dict
+    print("写入：configs/config.json")
+    with open("configs/config.json", "w") as f:
+        json.dump(config_template, f, indent=2)

preprocess_hubert_f0.py

@@ -0,0 +1,62 @@
+import math
+import multiprocessing
+import os
+import argparse
+from random import shuffle
+
+import paddle
+from glob import glob
+from tqdm import tqdm
+
+import utils
+import logging
+logging.getLogger('numba').setLevel(logging.WARNING)
+import librosa
+import numpy as np
+
+hps = utils.get_hparams_from_file("configs/config.json")
+sampling_rate = hps.data.sampling_rate
+hop_length = hps.data.hop_length
+
+
+def process_one(filename, hmodel):
+    # print(filename)
+    wav, sr = librosa.load(filename, sr=sampling_rate)
+    soft_path = filename + ".soft.pdtensor"
+    if not os.path.exists(soft_path):
+        devive = "cuda" if paddle.device.is_compiled_with_cuda() else "cpu"
+        wav16k = librosa.resample(wav, orig_sr=sampling_rate, target_sr=16000)
+        wav16k = paddle.to_tensor(wav16k).cpu() if devive=='cpu' else paddle.to_tensor(wav16k).cuda()
+        c:paddle.Tensor = utils.get_hubert_content(hmodel, wav_16k_tensor=wav16k)
+        paddle.save(c.cpu(), soft_path)
+    f0_path = filename + ".f0.npy"
+    if not os.path.exists(f0_path):
+        f0 = utils.compute_f0_dio(wav, sampling_rate=sampling_rate, hop_length=hop_length)
+        np.save(f0_path, f0)
+
+
+def process_batch(filenames):
+    print("正在加载内容的HuBERT……")
+    device = "cuda" if paddle.device.is_compiled_with_cuda() else "cpu"
+    hmodel = utils.get_hubert_model()
+    print("HuBERT已被装载。")
+    for filename in tqdm(filenames):
+        process_one(filename, hmodel)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--in_dir", type=str, default="dataset/44k", help="path to input dir")
+
+    args = parser.parse_args()
+    filenames = glob(f'{args.in_dir}/*/*.wav', recursive=True)  # [:10]
+    shuffle(filenames)
+    multiprocessing.set_start_method('spawn',force=True)
+
+    num_processes = 1
+    chunk_size = int(math.ceil(len(filenames) / num_processes))
+    chunks = [filenames[i:i + chunk_size] for i in range(0, len(filenames), chunk_size)]
+    print([len(c) for c in chunks])
+    processes = [multiprocessing.Process(target=process_batch, args=(chunk,)) for chunk in chunks]
+    for p in processes:
+        p.start()

raw/1.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5014db90a4bd53a18054ae2a2f9b1e733c8c474a6872beace1c3bd716b2cb61f
+size 2484268

requirements.txt

@@ -0,0 +1,16 @@
+numba==0.56.4 
+llvmlite==0.39.1 
+matplotlib 
+msgpack 
+librosa==0.10.0.post2 
+onnxruntime-gpu 
+pyworld==0.2.11.post0 
+praat-parselmouth 
+numpy==1.23.4 
+paddleaudio==1.0.2
+gradio==3.19.1
+pydub
+ffmpeg-python
+paddlepaddle==2.5.1
+visualdl
+tqdm