File size: 7,731 Bytes
9f38e54
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
import torch
import numpy as np
import re
import soundfile
import utils
import commons
import os
import librosa
from text import text_to_sequence
from mel_processing import spectrogram_torch
from models import SynthesizerTrn


class OpenVoiceBaseClass(object):
    def __init__(self, 
                config_path, 
                #device='cuda:0'):
                device="cpu"):
        #if 'cuda' in device:
        #    assert torch.cuda.is_available()

        hps = utils.get_hparams_from_file(config_path)

        model = SynthesizerTrn(
            len(getattr(hps, 'symbols', [])),
            hps.data.filter_length // 2 + 1,
            n_speakers=hps.data.n_speakers,
            **hps.model,
        ).to(device)

        model.eval()
        self.model = model
        self.hps = hps
        self.device = device

    def load_ckpt(self, ckpt_path):
        checkpoint_dict = torch.load(ckpt_path, map_location=torch.device('cpu'))
        a, b = self.model.load_state_dict(checkpoint_dict['model'], strict=False)
        print("Loaded checkpoint '{}'".format(ckpt_path))
        print('missing/unexpected keys:', a, b)


class BaseSpeakerTTS(OpenVoiceBaseClass):
    language_marks = {
        "english": "EN",
        "chinese": "ZH",
    }

    @staticmethod
    def get_text(text, hps, is_symbol):
        text_norm = text_to_sequence(text, hps.symbols, [] if is_symbol else hps.data.text_cleaners)
        if hps.data.add_blank:
            text_norm = commons.intersperse(text_norm, 0)
        text_norm = torch.LongTensor(text_norm)
        return text_norm

    @staticmethod
    def audio_numpy_concat(segment_data_list, sr, speed=1.):
        audio_segments = []
        for segment_data in segment_data_list:
            audio_segments += segment_data.reshape(-1).tolist()
            audio_segments += [0] * int((sr * 0.05)/speed)
        audio_segments = np.array(audio_segments).astype(np.float32)
        return audio_segments

    @staticmethod
    def split_sentences_into_pieces(text, language_str):
        texts = utils.split_sentence(text, language_str=language_str)
        print(" > Text splitted to sentences.")
        print('\n'.join(texts))
        print(" > ===========================")
        return texts

    def tts(self, text, output_path, speaker, language='English', speed=1.0):
        mark = self.language_marks.get(language.lower(), None)
        assert mark is not None, f"language {language} is not supported"

        texts = self.split_sentences_into_pieces(text, mark)

        audio_list = []
        for t in texts:
            t = re.sub(r'([a-z])([A-Z])', r'\1 \2', t)
            t = f'[{mark}]{t}[{mark}]'
            stn_tst = self.get_text(t, self.hps, False)
            device = self.device
            speaker_id = self.hps.speakers[speaker]
            with torch.no_grad():
                x_tst = stn_tst.unsqueeze(0).to(device)
                x_tst_lengths = torch.LongTensor([stn_tst.size(0)]).to(device)
                sid = torch.LongTensor([speaker_id]).to(device)
                audio = self.model.infer(x_tst, x_tst_lengths, sid=sid, noise_scale=0.667, noise_scale_w=0.6,
                                    length_scale=1.0 / speed)[0][0, 0].data.cpu().float().numpy()
            audio_list.append(audio)
        audio = self.audio_numpy_concat(audio_list, sr=self.hps.data.sampling_rate, speed=speed)

        if output_path is None:
            return audio
        else:
            soundfile.write(output_path, audio, self.hps.data.sampling_rate)


class ToneColorConverter(OpenVoiceBaseClass):
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)

        if kwargs.get('enable_watermark', True):
            import wavmark
            self.watermark_model = wavmark.load_model().to(self.device)
        else:
            self.watermark_model = None



    def extract_se(self, ref_wav_list, se_save_path=None):
        if isinstance(ref_wav_list, str):
            ref_wav_list = [ref_wav_list]
        
        device = self.device
        hps = self.hps
        gs = []
        
        for fname in ref_wav_list:
            audio_ref, sr = librosa.load(fname, sr=hps.data.sampling_rate)
            y = torch.FloatTensor(audio_ref)
            y = y.to(device)
            y = y.unsqueeze(0)
            y = spectrogram_torch(y, hps.data.filter_length,
                                        hps.data.sampling_rate, hps.data.hop_length, hps.data.win_length,
                                        center=False).to(device)
            with torch.no_grad():
                g = self.model.ref_enc(y.transpose(1, 2)).unsqueeze(-1)
                gs.append(g.detach())
        gs = torch.stack(gs).mean(0)

        if se_save_path is not None:
            os.makedirs(os.path.dirname(se_save_path), exist_ok=True)
            torch.save(gs.cpu(), se_save_path)

        return gs

    def convert(self, audio_src_path, src_se, tgt_se, output_path=None, tau=0.3, message="default"):
        hps = self.hps
        # load audio
        audio, sample_rate = librosa.load(audio_src_path, sr=hps.data.sampling_rate)
        audio = torch.tensor(audio).float()
        
        with torch.no_grad():
            y = torch.FloatTensor(audio).to(self.device)
            y = y.unsqueeze(0)
            spec = spectrogram_torch(y, hps.data.filter_length,
                                    hps.data.sampling_rate, hps.data.hop_length, hps.data.win_length,
                                    center=False).to(self.device)
            spec_lengths = torch.LongTensor([spec.size(-1)]).to(self.device)
            audio = self.model.voice_conversion(spec, spec_lengths, sid_src=src_se, sid_tgt=tgt_se, tau=tau)[0][
                        0, 0].data.cpu().float().numpy()
            audio = self.add_watermark(audio, message)
            if output_path is None:
                return audio
            else:
                soundfile.write(output_path, audio, hps.data.sampling_rate)
    
    def add_watermark(self, audio, message):
        if self.watermark_model is None:
            return audio
        device = self.device
        bits = utils.string_to_bits(message).reshape(-1)
        n_repeat = len(bits) // 32

        K = 16000
        coeff = 2
        for n in range(n_repeat):
            trunck = audio[(coeff * n) * K: (coeff * n + 1) * K]
            if len(trunck) != K:
                print('Audio too short, fail to add watermark')
                break
            message_npy = bits[n * 32: (n + 1) * 32]
            
            with torch.no_grad():
                signal = torch.FloatTensor(trunck).to(device)[None]
                message_tensor = torch.FloatTensor(message_npy).to(device)[None]
                signal_wmd_tensor = self.watermark_model.encode(signal, message_tensor)
                signal_wmd_npy = signal_wmd_tensor.detach().cpu().squeeze()
            audio[(coeff * n) * K: (coeff * n + 1) * K] = signal_wmd_npy
        return audio

    def detect_watermark(self, audio, n_repeat):
        bits = []
        K = 16000
        coeff = 2
        for n in range(n_repeat):
            trunck = audio[(coeff * n) * K: (coeff * n + 1) * K]
            if len(trunck) != K:
                print('Audio too short, fail to detect watermark')
                return 'Fail'
            with torch.no_grad():
                signal = torch.FloatTensor(trunck).to(self.device).unsqueeze(0)
                message_decoded_npy = (self.watermark_model.decode(signal) >= 0.5).int().detach().cpu().numpy().squeeze()
            bits.append(message_decoded_npy)
        bits = np.stack(bits).reshape(-1, 8)
        message = utils.bits_to_string(bits)
        return message