Spaces:
Sleeping
Sleeping
File size: 2,700 Bytes
f42d33e |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 |
from functools import lru_cache
import torch,json,os
import yaml
from scipy.io import wavfile
from mtts.text import TextProcessor
from mtts.models.fs2_model import FastSpeech2
import numpy as np
with open("dict_han_pinyin.json","r",encoding="utf-8") as f:
data_dict = json.load(f)
def normalize(wav):
assert wav.dtype == np.float32
eps = 1e-6
sil = wav[1500:2000]
#wav = wav - np.mean(sil)
#wav = (wav - np.min(wav))/(np.max(wav)-np.min(wav)+eps)
wav = wav / np.max(np.abs(wav))
#wav = wav*2-1
wav = wav * 32767
return wav.astype('int16')
def to_int16(wav):
wav = wav = wav * 32767
wav = np.clamp(wav, -32767, 32768)
return wav.astype('int16')
def __build_vocoder(config):
vocoder_name = config['vocoder']['type']
VocoderClass = eval(vocoder_name)
model = VocoderClass(config=config['vocoder'][vocoder_name])
return model
@lru_cache(maxsize=10)
def get_pretrained_model(line):
config = "text_to_speech\examples/biaobei\config.yaml"
checkpoint = "text_to_speech\checkpoints\checkpoint_140000.pth.tar"
with open(config) as f:
config = yaml.safe_load(f)
sr = config['fbank']['sample_rate']
vocoder = __build_vocoder(config)
text_processor = TextProcessor(config)
model = FastSpeech2(config)
if checkpoint != '':
sd = torch.load(checkpoint, map_location="cpu")
if 'model' in sd.keys():
sd = sd['model']
model.load_state_dict(sd)
del sd # to save mem
model = model.to("cpu")
torch.set_grad_enabled(False)
pinyin = ""
hanzi = ""
for i in line:
pinyin+=data_dict[i]+" "
hanzi +=i+" "
post_line = f"text1|sil {pinyin}sil|sil {hanzi}sil|0"
name, tokens = text_processor(post_line)
tokens = tokens.to("cpu")
seq_len = torch.tensor([tokens.shape[1]])
tokens = tokens.unsqueeze(1)
seq_len = seq_len.to("cpu")
max_src_len = torch.max(seq_len)
output = model(tokens, seq_len, max_src_len=max_src_len, d_control=1.0)
mel_pred, mel_postnet, d_pred, src_mask, mel_mask, mel_len = output
# convert to waveform using vocoder
mel_postnet = mel_postnet[0].transpose(0, 1).detach()
mel_postnet += config['fbank']['mel_mean']
wav = vocoder(mel_postnet)
if config['synthesis']['normalize']:
wav = normalize(wav)
else:
wav = to_int16(wav)
dst_file = os.path.join(f'{name}.wav')
#np.save(dst_file+'.npy',mel_postnet.cpu().numpy())
wavfile.write(dst_file, sr, wav)
return dst_file,2.0
chinese_models = {
"csukuangfj/vits-piper-zh_CN-huayan-medium": 1}
language_to_models = {
"Chinese (Mandarin, 普通话)": list(chinese_models.keys())} |