soiz1 commited on
Commit
91646d2
·
verified ·
1 Parent(s): 538cd32

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +739 -0
app.py CHANGED
@@ -1,4 +1,743 @@
 
 
1
  import gradio as gr
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2
 
3
  gallery_data = {"sikokumetan": {"webp": "default/sikokumetan.webp", "mp3": "default/sikokumetan.mp3"}}
4
 
 
1
+ import os
2
+ import spaces
3
  import gradio as gr
4
+ import torch
5
+ import torchaudio
6
+ import librosa
7
+ from modules.commons import build_model, load_checkpoint, recursive_munch
8
+ import yaml
9
+ from hf_utils import load_custom_model_from_hf
10
+ import numpy as np
11
+ from pydub import AudioSegment
12
+
13
+ # Load model and configuration
14
+ device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
15
+
16
+ dit_checkpoint_path, dit_config_path = load_custom_model_from_hf("Plachta/Seed-VC",
17
+ "DiT_seed_v2_uvit_whisper_small_wavenet_bigvgan_pruned.pth",
18
+ "config_dit_mel_seed_uvit_whisper_small_wavenet.yml")
19
+ # dit_checkpoint_path = "E:/DiT_epoch_00018_step_801000.pth"
20
+ # dit_config_path = "configs/config_dit_mel_seed_uvit_whisper_small_encoder_wavenet.yml"
21
+ config = yaml.safe_load(open(dit_config_path, 'r'))
22
+ model_params = recursive_munch(config['model_params'])
23
+ model = build_model(model_params, stage='DiT')
24
+ hop_length = config['preprocess_params']['spect_params']['hop_length']
25
+ sr = config['preprocess_params']['sr']
26
+
27
+ # Load checkpoints
28
+ model, _, _, _ = load_checkpoint(model, None, dit_checkpoint_path,
29
+ load_only_params=True, ignore_modules=[], is_distributed=False)
30
+ for key in model:
31
+ model[key].eval()
32
+ model[key].to(device)
33
+ model.cfm.estimator.setup_caches(max_batch_size=1, max_seq_length=8192)
34
+
35
+ # Load additional modules
36
+ from modules.campplus.DTDNN import CAMPPlus
37
+
38
+ campplus_ckpt_path = load_custom_model_from_hf("funasr/campplus", "campplus_cn_common.bin", config_filename=None)
39
+ campplus_model = CAMPPlus(feat_dim=80, embedding_size=192)
40
+ campplus_model.load_state_dict(torch.load(campplus_ckpt_path, map_location="cpu"))
41
+ campplus_model.eval()
42
+ campplus_model.to(device)
43
+
44
+ from modules.bigvgan import bigvgan
45
+
46
+ bigvgan_model = bigvgan.BigVGAN.from_pretrained('nvidia/bigvgan_v2_22khz_80band_256x', use_cuda_kernel=False)
47
+
48
+ # remove weight norm in the model and set to eval mode
49
+ bigvgan_model.remove_weight_norm()
50
+ bigvgan_model = bigvgan_model.eval().to(device)
51
+
52
+ ckpt_path, config_path = load_custom_model_from_hf("Plachta/FAcodec", 'pytorch_model.bin', 'config.yml')
53
+
54
+ codec_config = yaml.safe_load(open(config_path))
55
+ codec_model_params = recursive_munch(codec_config['model_params'])
56
+ codec_encoder = build_model(codec_model_params, stage="codec")
57
+
58
+ ckpt_params = torch.load(ckpt_path, map_location="cpu")
59
+
60
+ for key in codec_encoder:
61
+ codec_encoder[key].load_state_dict(ckpt_params[key], strict=False)
62
+ _ = [codec_encoder[key].eval() for key in codec_encoder]
63
+ _ = [codec_encoder[key].to(device) for key in codec_encoder]
64
+
65
+ # whisper
66
+ from transformers import AutoFeatureExtractor, WhisperModel
67
+
68
+ whisper_name = model_params.speech_tokenizer.whisper_name if hasattr(model_params.speech_tokenizer,
69
+ 'whisper_name') else "openai/whisper-small"
70
+ whisper_model = WhisperModel.from_pretrained(whisper_name, torch_dtype=torch.float16).to(device)
71
+ del whisper_model.decoder
72
+ whisper_feature_extractor = AutoFeatureExtractor.from_pretrained(whisper_name)
73
+
74
+ # Generate mel spectrograms
75
+ mel_fn_args = {
76
+ "n_fft": config['preprocess_params']['spect_params']['n_fft'],
77
+ "win_size": config['preprocess_params']['spect_params']['win_length'],
78
+ "hop_size": config['preprocess_params']['spect_params']['hop_length'],
79
+ "num_mels": config['preprocess_params']['spect_params']['n_mels'],
80
+ "sampling_rate": sr,
81
+ "fmin": 0,
82
+ "fmax": None,
83
+ "center": False
84
+ }
85
+ from modules.audio import mel_spectrogram
86
+
87
+ to_mel = lambda x: mel_spectrogram(x, **mel_fn_args)
88
+
89
+ # f0 conditioned model
90
+ dit_checkpoint_path, dit_config_path = load_custom_model_from_hf("Plachta/Seed-VC",
91
+ "DiT_seed_v2_uvit_whisper_base_f0_44k_bigvgan_pruned_ft_ema.pth",
92
+ "config_dit_mel_seed_uvit_whisper_base_f0_44k.yml")
93
+
94
+ config = yaml.safe_load(open(dit_config_path, 'r'))
95
+ model_params = recursive_munch(config['model_params'])
96
+ model_f0 = build_model(model_params, stage='DiT')
97
+ hop_length = config['preprocess_params']['spect_params']['hop_length']
98
+ sr = config['preprocess_params']['sr']
99
+
100
+ # Load checkpoints
101
+ model_f0, _, _, _ = load_checkpoint(model_f0, None, dit_checkpoint_path,
102
+ load_only_params=True, ignore_modules=[], is_distributed=False)
103
+ for key in model_f0:
104
+ model_f0[key].eval()
105
+ model_f0[key].to(device)
106
+ model_f0.cfm.estimator.setup_caches(max_batch_size=1, max_seq_length=8192)
107
+
108
+ # f0 extractor
109
+ from modules.rmvpe import RMVPE
110
+
111
+ model_path = load_custom_model_from_hf("lj1995/VoiceConversionWebUI", "rmvpe.pt", None)
112
+ rmvpe = RMVPE(model_path, is_half=False, device=device)
113
+
114
+ mel_fn_args_f0 = {
115
+ "n_fft": config['preprocess_params']['spect_params']['n_fft'],
116
+ "win_size": config['preprocess_params']['spect_params']['win_length'],
117
+ "hop_size": config['preprocess_params']['spect_params']['hop_length'],
118
+ "num_mels": config['preprocess_params']['spect_params']['n_mels'],
119
+ "sampling_rate": sr,
120
+ "fmin": 0,
121
+ "fmax": None,
122
+ "center": False
123
+ }
124
+ to_mel_f0 = lambda x: mel_spectrogram(x, **mel_fn_args_f0)
125
+ bigvgan_44k_model = bigvgan.BigVGAN.from_pretrained('nvidia/bigvgan_v2_44khz_128band_512x', use_cuda_kernel=False)
126
+
127
+ # remove weight norm in the model and set to eval mode
128
+ bigvgan_44k_model.remove_weight_norm()
129
+ bigvgan_44k_model = bigvgan_44k_model.eval().to(device)
130
+
131
+ def adjust_f0_semitones(f0_sequence, n_semitones):
132
+ factor = 2 ** (n_semitones / 12)
133
+ return f0_sequence * factor
134
+
135
+ def crossfade(chunk1, chunk2, overlap):
136
+ fade_out = np.cos(np.linspace(0, np.pi / 2, overlap)) ** 2
137
+ fade_in = np.cos(np.linspace(np.pi / 2, 0, overlap)) ** 2
138
+ chunk2[:overlap] = chunk2[:overlap] * fade_in + chunk1[-overlap:] * fade_out
139
+ return chunk2
140
+
141
+ # streaming and chunk processing related params
142
+ bitrate = "320k"
143
+ overlap_frame_len = 16
144
+ @spaces.GPU
145
+ @torch.no_grad()
146
+ @torch.inference_mode()
147
+ def voice_conversion(source, target, diffusion_steps, length_adjust, inference_cfg_rate, f0_condition, auto_f0_adjust, pitch_shift):
148
+ inference_module = model if not f0_condition else model_f0
149
+ mel_fn = to_mel if not f0_condition else to_mel_f0
150
+ bigvgan_fn = bigvgan_model if not f0_condition else bigvgan_44k_model
151
+ sr = 22050 if not f0_condition else 44100
152
+ hop_length = 256 if not f0_condition else 512
153
+ max_context_window = sr // hop_length * 30
154
+ overlap_wave_len = overlap_frame_len * hop_length
155
+ # Load audio
156
+ source_audio = librosa.load(source, sr=sr)[0]
157
+ ref_audio = librosa.load(target, sr=sr)[0]
158
+
159
+ # Process audio
160
+ source_audio = torch.tensor(source_audio).unsqueeze(0).float().to(device)
161
+ ref_audio = torch.tensor(ref_audio[:sr * 25]).unsqueeze(0).float().to(device)
162
+
163
+ # Resample
164
+ ref_waves_16k = torchaudio.functional.resample(ref_audio, sr, 16000)
165
+ converted_waves_16k = torchaudio.functional.resample(source_audio, sr, 16000)
166
+ # if source audio less than 30 seconds, whisper can handle in one forward
167
+ if converted_waves_16k.size(-1) <= 16000 * 30:
168
+ alt_inputs = whisper_feature_extractor([converted_waves_16k.squeeze(0).cpu().numpy()],
169
+ return_tensors="pt",
170
+ return_attention_mask=True,
171
+ sampling_rate=16000)
172
+ alt_input_features = whisper_model._mask_input_features(
173
+ alt_inputs.input_features, attention_mask=alt_inputs.attention_mask).to(device)
174
+ alt_outputs = whisper_model.encoder(
175
+ alt_input_features.to(whisper_model.encoder.dtype),
176
+ head_mask=None,
177
+ output_attentions=False,
178
+ output_hidden_states=False,
179
+ return_dict=True,
180
+ )
181
+ S_alt = alt_outputs.last_hidden_state.to(torch.float32)
182
+ S_alt = S_alt[:, :converted_waves_16k.size(-1) // 320 + 1]
183
+ else:
184
+ overlapping_time = 5 # 5 seconds
185
+ S_alt_list = []
186
+ buffer = None
187
+ traversed_time = 0
188
+ while traversed_time < converted_waves_16k.size(-1):
189
+ if buffer is None: # first chunk
190
+ chunk = converted_waves_16k[:, traversed_time:traversed_time + 16000 * 30]
191
+ else:
192
+ chunk = torch.cat([buffer, converted_waves_16k[:, traversed_time:traversed_time + 16000 * (30 - overlapping_time)]], dim=-1)
193
+ alt_inputs = whisper_feature_extractor([chunk.squeeze(0).cpu().numpy()],
194
+ return_tensors="pt",
195
+ return_attention_mask=True,
196
+ sampling_rate=16000)
197
+ alt_input_features = whisper_model._mask_input_features(
198
+ alt_inputs.input_features, attention_mask=alt_inputs.attention_mask).to(device)
199
+ alt_outputs = whisper_model.encoder(
200
+ alt_input_features.to(whisper_model.encoder.dtype),
201
+ head_mask=None,
202
+ output_attentions=False,
203
+ output_hidden_states=False,
204
+ return_dict=True,
205
+ )
206
+ S_alt = alt_outputs.last_hidden_state.to(torch.float32)
207
+ S_alt = S_alt[:, :chunk.size(-1) // 320 + 1]
208
+ if traversed_time == 0:
209
+ S_alt_list.append(S_alt)
210
+ else:
211
+ S_alt_list.append(S_alt[:, 50 * overlapping_time:])
212
+ buffer = chunk[:, -16000 * overlapping_time:]
213
+ traversed_time += 30 * 16000 if traversed_time == 0 else chunk.size(-1) - 16000 * overlapping_time
214
+ S_alt = torch.cat(S_alt_list, dim=1)
215
+
216
+ ori_waves_16k = torchaudio.functional.resample(ref_audio, sr, 16000)
217
+ ori_inputs = whisper_feature_extractor([ori_waves_16k.squeeze(0).cpu().numpy()],
218
+ return_tensors="pt",
219
+ return_attention_mask=True)
220
+ ori_input_features = whisper_model._mask_input_features(
221
+ ori_inputs.input_features, attention_mask=ori_inputs.attention_mask).to(device)
222
+ with torch.no_grad():
223
+ ori_outputs = whisper_model.encoder(
224
+ ori_input_features.to(whisper_model.encoder.dtype),
225
+ head_mask=None,
226
+ output_attentions=False,
227
+ output_hidden_states=False,
228
+ return_dict=True,
229
+ )
230
+ S_ori = ori_outputs.last_hidden_state.to(torch.float32)
231
+ S_ori = S_ori[:, :ori_waves_16k.size(-1) // 320 + 1]
232
+
233
+ mel = mel_fn(source_audio.to(device).float())
234
+ mel2 = mel_fn(ref_audio.to(device).float())
235
+
236
+ target_lengths = torch.LongTensor([int(mel.size(2) * length_adjust)]).to(mel.device)
237
+ target2_lengths = torch.LongTensor([mel2.size(2)]).to(mel2.device)
238
+
239
+ feat2 = torchaudio.compliance.kaldi.fbank(ref_waves_16k,
240
+ num_mel_bins=80,
241
+ dither=0,
242
+ sample_frequency=16000)
243
+ feat2 = feat2 - feat2.mean(dim=0, keepdim=True)
244
+ style2 = campplus_model(feat2.unsqueeze(0))
245
+
246
+ if f0_condition:
247
+ F0_ori = rmvpe.infer_from_audio(ref_waves_16k[0], thred=0.5)
248
+ F0_alt = rmvpe.infer_from_audio(converted_waves_16k[0], thred=0.5)
249
+
250
+ F0_ori = torch.from_numpy(F0_ori).to(device)[None]
251
+ F0_alt = torch.from_numpy(F0_alt).to(device)[None]
252
+
253
+ voiced_F0_ori = F0_ori[F0_ori > 1]
254
+ voiced_F0_alt = F0_alt[F0_alt > 1]
255
+
256
+ log_f0_alt = torch.log(F0_alt + 1e-5)
257
+ voiced_log_f0_ori = torch.log(voiced_F0_ori + 1e-5)
258
+ voiced_log_f0_alt = torch.log(voiced_F0_alt + 1e-5)
259
+ median_log_f0_ori = torch.median(voiced_log_f0_ori)
260
+ median_log_f0_alt = torch.median(voiced_log_f0_alt)
261
+
262
+ # shift alt log f0 level to ori log f0 level
263
+ shifted_log_f0_alt = log_f0_alt.clone()
264
+ if auto_f0_adjust:
265
+ shifted_log_f0_alt[F0_alt > 1] = log_f0_alt[F0_alt > 1] - median_log_f0_alt + median_log_f0_ori
266
+ shifted_f0_alt = torch.exp(shifted_log_f0_alt)
267
+ if pitch_shift != 0:
268
+ shifted_f0_alt[F0_alt > 1] = adjust_f0_semitones(shifted_f0_alt[F0_alt > 1], pitch_shift)
269
+ else:
270
+ F0_ori = None
271
+ F0_alt = None
272
+ shifted_f0_alt = None
273
+
274
+ # Length regulation
275
+ cond, _, codes, commitment_loss, codebook_loss = inference_module.length_regulator(S_alt, ylens=target_lengths, n_quantizers=3, f0=shifted_f0_alt)
276
+ prompt_condition, _, codes, commitment_loss, codebook_loss = inference_module.length_regulator(S_ori, ylens=target2_lengths, n_quantizers=3, f0=F0_ori)
277
+
278
+ max_source_window = max_context_window - mel2.size(2)
279
+ # split source condition (cond) into chunks
280
+ processed_frames = 0
281
+ generated_wave_chunks = []
282
+ # generate chunk by chunk and stream the output
283
+ while processed_frames < cond.size(1):
284
+ chunk_cond = cond[:, processed_frames:processed_frames + max_source_window]
285
+ is_last_chunk = processed_frames + max_source_window >= cond.size(1)
286
+ cat_condition = torch.cat([prompt_condition, chunk_cond], dim=1)
287
+ with torch.autocast(device_type='cuda', dtype=torch.float16):
288
+ # Voice Conversion
289
+ vc_target = inference_module.cfm.inference(cat_condition,
290
+ torch.LongTensor([cat_condition.size(1)]).to(mel2.device),
291
+ mel2, style2, None, diffusion_steps,
292
+ inference_cfg_rate=inference_cfg_rate)
293
+ vc_target = vc_target[:, :, mel2.size(-1):]
294
+ vc_wave = bigvgan_fn(vc_target.float())[0]
295
+ if processed_frames == 0:
296
+ if is_last_chunk:
297
+ output_wave = vc_wave[0].cpu().numpy()
298
+ generated_wave_chunks.append(output_wave)
299
+ output_wave = (output_wave * 32768.0).astype(np.int16)
300
+ mp3_bytes = AudioSegment(
301
+ output_wave.tobytes(), frame_rate=sr,
302
+ sample_width=output_wave.dtype.itemsize, channels=1
303
+ ).export(format="mp3", bitrate=bitrate).read()
304
+ yield mp3_bytes, (sr, np.concatenate(generated_wave_chunks))
305
+ break
306
+ output_wave = vc_wave[0, :-overlap_wave_len].cpu().numpy()
307
+ generated_wave_chunks.append(output_wave)
308
+ previous_chunk = vc_wave[0, -overlap_wave_len:]
309
+ processed_frames += vc_target.size(2) - overlap_frame_len
310
+ output_wave = (output_wave * 32768.0).astype(np.int16)
311
+ mp3_bytes = AudioSegment(
312
+ output_wave.tobytes(), frame_rate=sr,
313
+ sample_width=output_wave.dtype.itemsize, channels=1
314
+ ).export(format="mp3", bitrate=bitrate).read()
315
+ yield mp3_bytes, None
316
+ elif is_last_chunk:
317
+ output_wave = crossfade(previous_chunk.cpu().numpy(), vc_wave[0].cpu().numpy(), overlap_wave_len)
318
+ generated_wave_chunks.append(output_wave)
319
+ processed_frames += vc_target.size(2) - overlap_frame_len
320
+ output_wave = (output_wave * 32768.0).astype(np.int16)
321
+ mp3_bytes = AudioSegment(
322
+ output_wave.tobytes(), frame_rate=sr,
323
+ sample_width=output_wave.dtype.itemsize, channels=1
324
+ ).export(format="mp3", bitrate=bitrate).read()
325
+ yield mp3_bytes, (sr, np.concatenate(generated_wave_chunks))
326
+ break
327
+ else:
328
+ output_wave = crossfade(previous_chunk.cpu().numpy(), vc_wave[0, :-overlap_wave_len].cpu().numpy(), overlap_wave_len)
329
+ generated_wave_chunks.append(output_wave)
330
+ previous_chunk = vc_wave[0, -overlap_wave_len:]
331
+ processed_frames += vc_target.size(2) - overlap_frame_len
332
+ output_wave = (output_wave * 32768.0).astype(np.int16)
333
+ mp3_bytes = AudioSegment(
334
+ output_wave.tobytes(), frame_rate=sr,
335
+ sample_width=output_wave.dtype.itemsize, channels=1
336
+ ).export(format="mp3", bitrate=bitrate).read()
337
+ yield mp3_bytes, None
338
+
339
+ import os
340
+ import spaces
341
+ import gradio as gr
342
+ import torch
343
+ import torchaudio
344
+ import librosa
345
+ from modules.commons import build_model, load_checkpoint, recursive_munch
346
+ import yaml
347
+ from hf_utils import load_custom_model_from_hf
348
+ import numpy as np
349
+ from pydub import AudioSegment
350
+
351
+ # Load model and configuration
352
+ device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
353
+
354
+ dit_checkpoint_path, dit_config_path = load_custom_model_from_hf("Plachta/Seed-VC",
355
+ "DiT_seed_v2_uvit_whisper_small_wavenet_bigvgan_pruned.pth",
356
+ "config_dit_mel_seed_uvit_whisper_small_wavenet.yml")
357
+ # dit_checkpoint_path = "E:/DiT_epoch_00018_step_801000.pth"
358
+ # dit_config_path = "configs/config_dit_mel_seed_uvit_whisper_small_encoder_wavenet.yml"
359
+ config = yaml.safe_load(open(dit_config_path, 'r'))
360
+ model_params = recursive_munch(config['model_params'])
361
+ model = build_model(model_params, stage='DiT')
362
+ hop_length = config['preprocess_params']['spect_params']['hop_length']
363
+ sr = config['preprocess_params']['sr']
364
+
365
+ # Load checkpoints
366
+ model, _, _, _ = load_checkpoint(model, None, dit_checkpoint_path,
367
+ load_only_params=True, ignore_modules=[], is_distributed=False)
368
+ for key in model:
369
+ model[key].eval()
370
+ model[key].to(device)
371
+ model.cfm.estimator.setup_caches(max_batch_size=1, max_seq_length=8192)
372
+
373
+ # Load additional modules
374
+ from modules.campplus.DTDNN import CAMPPlus
375
+
376
+ campplus_ckpt_path = load_custom_model_from_hf("funasr/campplus", "campplus_cn_common.bin", config_filename=None)
377
+ campplus_model = CAMPPlus(feat_dim=80, embedding_size=192)
378
+ campplus_model.load_state_dict(torch.load(campplus_ckpt_path, map_location="cpu"))
379
+ campplus_model.eval()
380
+ campplus_model.to(device)
381
+
382
+ from modules.bigvgan import bigvgan
383
+
384
+ bigvgan_model = bigvgan.BigVGAN.from_pretrained('nvidia/bigvgan_v2_22khz_80band_256x', use_cuda_kernel=False)
385
+
386
+ # remove weight norm in the model and set to eval mode
387
+ bigvgan_model.remove_weight_norm()
388
+ bigvgan_model = bigvgan_model.eval().to(device)
389
+
390
+ ckpt_path, config_path = load_custom_model_from_hf("Plachta/FAcodec", 'pytorch_model.bin', 'config.yml')
391
+
392
+ codec_config = yaml.safe_load(open(config_path))
393
+ codec_model_params = recursive_munch(codec_config['model_params'])
394
+ codec_encoder = build_model(codec_model_params, stage="codec")
395
+
396
+ ckpt_params = torch.load(ckpt_path, map_location="cpu")
397
+
398
+ for key in codec_encoder:
399
+ codec_encoder[key].load_state_dict(ckpt_params[key], strict=False)
400
+ _ = [codec_encoder[key].eval() for key in codec_encoder]
401
+ _ = [codec_encoder[key].to(device) for key in codec_encoder]
402
+
403
+ # whisper
404
+ from transformers import AutoFeatureExtractor, WhisperModel
405
+
406
+ whisper_name = model_params.speech_tokenizer.whisper_name if hasattr(model_params.speech_tokenizer,
407
+ 'whisper_name') else "openai/whisper-small"
408
+ whisper_model = WhisperModel.from_pretrained(whisper_name, torch_dtype=torch.float16).to(device)
409
+ del whisper_model.decoder
410
+ whisper_feature_extractor = AutoFeatureExtractor.from_pretrained(whisper_name)
411
+
412
+ # Generate mel spectrograms
413
+ mel_fn_args = {
414
+ "n_fft": config['preprocess_params']['spect_params']['n_fft'],
415
+ "win_size": config['preprocess_params']['spect_params']['win_length'],
416
+ "hop_size": config['preprocess_params']['spect_params']['hop_length'],
417
+ "num_mels": config['preprocess_params']['spect_params']['n_mels'],
418
+ "sampling_rate": sr,
419
+ "fmin": 0,
420
+ "fmax": None,
421
+ "center": False
422
+ }
423
+ from modules.audio import mel_spectrogram
424
+
425
+ to_mel = lambda x: mel_spectrogram(x, **mel_fn_args)
426
+
427
+ # f0 conditioned model
428
+ dit_checkpoint_path, dit_config_path = load_custom_model_from_hf("Plachta/Seed-VC",
429
+ "DiT_seed_v2_uvit_whisper_base_f0_44k_bigvgan_pruned_ft_ema.pth",
430
+ "config_dit_mel_seed_uvit_whisper_base_f0_44k.yml")
431
+
432
+ config = yaml.safe_load(open(dit_config_path, 'r'))
433
+ model_params = recursive_munch(config['model_params'])
434
+ model_f0 = build_model(model_params, stage='DiT')
435
+ hop_length = config['preprocess_params']['spect_params']['hop_length']
436
+ sr = config['preprocess_params']['sr']
437
+
438
+ # Load checkpoints
439
+ model_f0, _, _, _ = load_checkpoint(model_f0, None, dit_checkpoint_path,
440
+ load_only_params=True, ignore_modules=[], is_distributed=False)
441
+ for key in model_f0:
442
+ model_f0[key].eval()
443
+ model_f0[key].to(device)
444
+ model_f0.cfm.estimator.setup_caches(max_batch_size=1, max_seq_length=8192)
445
+
446
+ # f0 extractor
447
+ from modules.rmvpe import RMVPE
448
+
449
+ model_path = load_custom_model_from_hf("lj1995/VoiceConversionWebUI", "rmvpe.pt", None)
450
+ rmvpe = RMVPE(model_path, is_half=False, device=device)
451
+
452
+ mel_fn_args_f0 = {
453
+ "n_fft": config['preprocess_params']['spect_params']['n_fft'],
454
+ "win_size": config['preprocess_params']['spect_params']['win_length'],
455
+ "hop_size": config['preprocess_params']['spect_params']['hop_length'],
456
+ "num_mels": config['preprocess_params']['spect_params']['n_mels'],
457
+ "sampling_rate": sr,
458
+ "fmin": 0,
459
+ "fmax": None,
460
+ "center": False
461
+ }
462
+ to_mel_f0 = lambda x: mel_spectrogram(x, **mel_fn_args_f0)
463
+ bigvgan_44k_model = bigvgan.BigVGAN.from_pretrained('nvidia/bigvgan_v2_44khz_128band_512x', use_cuda_kernel=False)
464
+
465
+ # remove weight norm in the model and set to eval mode
466
+ bigvgan_44k_model.remove_weight_norm()
467
+ bigvgan_44k_model = bigvgan_44k_model.eval().to(device)
468
+
469
+ def adjust_f0_semitones(f0_sequence, n_semitones):
470
+ factor = 2 ** (n_semitones / 12)
471
+ return f0_sequence * factor
472
+
473
+ def crossfade(chunk1, chunk2, overlap):
474
+ fade_out = np.cos(np.linspace(0, np.pi / 2, overlap)) ** 2
475
+ fade_in = np.cos(np.linspace(np.pi / 2, 0, overlap)) ** 2
476
+ chunk2[:overlap] = chunk2[:overlap] * fade_in + chunk1[-overlap:] * fade_out
477
+ return chunk2
478
+
479
+ # streaming and chunk processing related params
480
+ bitrate = "320k"
481
+ overlap_frame_len = 16
482
+ @spaces.GPU
483
+ @torch.no_grad()
484
+ @torch.inference_mode()
485
+ def voice_conversion(source, target, diffusion_steps, length_adjust, inference_cfg_rate, f0_condition, auto_f0_adjust, pitch_shift):
486
+ inference_module = model if not f0_condition else model_f0
487
+ mel_fn = to_mel if not f0_condition else to_mel_f0
488
+ bigvgan_fn = bigvgan_model if not f0_condition else bigvgan_44k_model
489
+ sr = 22050 if not f0_condition else 44100
490
+ hop_length = 256 if not f0_condition else 512
491
+ max_context_window = sr // hop_length * 30
492
+ overlap_wave_len = overlap_frame_len * hop_length
493
+ # Load audio
494
+ source_audio = librosa.load(source, sr=sr)[0]
495
+ ref_audio = librosa.load(target, sr=sr)[0]
496
+
497
+ # Process audio
498
+ source_audio = torch.tensor(source_audio).unsqueeze(0).float().to(device)
499
+ ref_audio = torch.tensor(ref_audio[:sr * 25]).unsqueeze(0).float().to(device)
500
+
501
+ # Resample
502
+ ref_waves_16k = torchaudio.functional.resample(ref_audio, sr, 16000)
503
+ converted_waves_16k = torchaudio.functional.resample(source_audio, sr, 16000)
504
+ # if source audio less than 30 seconds, whisper can handle in one forward
505
+ if converted_waves_16k.size(-1) <= 16000 * 30:
506
+ alt_inputs = whisper_feature_extractor([converted_waves_16k.squeeze(0).cpu().numpy()],
507
+ return_tensors="pt",
508
+ return_attention_mask=True,
509
+ sampling_rate=16000)
510
+ alt_input_features = whisper_model._mask_input_features(
511
+ alt_inputs.input_features, attention_mask=alt_inputs.attention_mask).to(device)
512
+ alt_outputs = whisper_model.encoder(
513
+ alt_input_features.to(whisper_model.encoder.dtype),
514
+ head_mask=None,
515
+ output_attentions=False,
516
+ output_hidden_states=False,
517
+ return_dict=True,
518
+ )
519
+ S_alt = alt_outputs.last_hidden_state.to(torch.float32)
520
+ S_alt = S_alt[:, :converted_waves_16k.size(-1) // 320 + 1]
521
+ else:
522
+ overlapping_time = 5 # 5 seconds
523
+ S_alt_list = []
524
+ buffer = None
525
+ traversed_time = 0
526
+ while traversed_time < converted_waves_16k.size(-1):
527
+ if buffer is None: # first chunk
528
+ chunk = converted_waves_16k[:, traversed_time:traversed_time + 16000 * 30]
529
+ else:
530
+ chunk = torch.cat([buffer, converted_waves_16k[:, traversed_time:traversed_time + 16000 * (30 - overlapping_time)]], dim=-1)
531
+ alt_inputs = whisper_feature_extractor([chunk.squeeze(0).cpu().numpy()],
532
+ return_tensors="pt",
533
+ return_attention_mask=True,
534
+ sampling_rate=16000)
535
+ alt_input_features = whisper_model._mask_input_features(
536
+ alt_inputs.input_features, attention_mask=alt_inputs.attention_mask).to(device)
537
+ alt_outputs = whisper_model.encoder(
538
+ alt_input_features.to(whisper_model.encoder.dtype),
539
+ head_mask=None,
540
+ output_attentions=False,
541
+ output_hidden_states=False,
542
+ return_dict=True,
543
+ )
544
+ S_alt = alt_outputs.last_hidden_state.to(torch.float32)
545
+ S_alt = S_alt[:, :chunk.size(-1) // 320 + 1]
546
+ if traversed_time == 0:
547
+ S_alt_list.append(S_alt)
548
+ else:
549
+ S_alt_list.append(S_alt[:, 50 * overlapping_time:])
550
+ buffer = chunk[:, -16000 * overlapping_time:]
551
+ traversed_time += 30 * 16000 if traversed_time == 0 else chunk.size(-1) - 16000 * overlapping_time
552
+ S_alt = torch.cat(S_alt_list, dim=1)
553
+
554
+ ori_waves_16k = torchaudio.functional.resample(ref_audio, sr, 16000)
555
+ ori_inputs = whisper_feature_extractor([ori_waves_16k.squeeze(0).cpu().numpy()],
556
+ return_tensors="pt",
557
+ return_attention_mask=True)
558
+ ori_input_features = whisper_model._mask_input_features(
559
+ ori_inputs.input_features, attention_mask=ori_inputs.attention_mask).to(device)
560
+ with torch.no_grad():
561
+ ori_outputs = whisper_model.encoder(
562
+ ori_input_features.to(whisper_model.encoder.dtype),
563
+ head_mask=None,
564
+ output_attentions=False,
565
+ output_hidden_states=False,
566
+ return_dict=True,
567
+ )
568
+ S_ori = ori_outputs.last_hidden_state.to(torch.float32)
569
+ S_ori = S_ori[:, :ori_waves_16k.size(-1) // 320 + 1]
570
+
571
+ mel = mel_fn(source_audio.to(device).float())
572
+ mel2 = mel_fn(ref_audio.to(device).float())
573
+
574
+ target_lengths = torch.LongTensor([int(mel.size(2) * length_adjust)]).to(mel.device)
575
+ target2_lengths = torch.LongTensor([mel2.size(2)]).to(mel2.device)
576
+
577
+ feat2 = torchaudio.compliance.kaldi.fbank(ref_waves_16k,
578
+ num_mel_bins=80,
579
+ dither=0,
580
+ sample_frequency=16000)
581
+ feat2 = feat2 - feat2.mean(dim=0, keepdim=True)
582
+ style2 = campplus_model(feat2.unsqueeze(0))
583
+
584
+ if f0_condition:
585
+ F0_ori = rmvpe.infer_from_audio(ref_waves_16k[0], thred=0.5)
586
+ F0_alt = rmvpe.infer_from_audio(converted_waves_16k[0], thred=0.5)
587
+
588
+ F0_ori = torch.from_numpy(F0_ori).to(device)[None]
589
+ F0_alt = torch.from_numpy(F0_alt).to(device)[None]
590
+
591
+ voiced_F0_ori = F0_ori[F0_ori > 1]
592
+ voiced_F0_alt = F0_alt[F0_alt > 1]
593
+
594
+ log_f0_alt = torch.log(F0_alt + 1e-5)
595
+ voiced_log_f0_ori = torch.log(voiced_F0_ori + 1e-5)
596
+ voiced_log_f0_alt = torch.log(voiced_F0_alt + 1e-5)
597
+ median_log_f0_ori = torch.median(voiced_log_f0_ori)
598
+ median_log_f0_alt = torch.median(voiced_log_f0_alt)
599
+
600
+ # shift alt log f0 level to ori log f0 level
601
+ shifted_log_f0_alt = log_f0_alt.clone()
602
+ if auto_f0_adjust:
603
+ shifted_log_f0_alt[F0_alt > 1] = log_f0_alt[F0_alt > 1] - median_log_f0_alt + median_log_f0_ori
604
+ shifted_f0_alt = torch.exp(shifted_log_f0_alt)
605
+ if pitch_shift != 0:
606
+ shifted_f0_alt[F0_alt > 1] = adjust_f0_semitones(shifted_f0_alt[F0_alt > 1], pitch_shift)
607
+ else:
608
+ F0_ori = None
609
+ F0_alt = None
610
+ shifted_f0_alt = None
611
+
612
+ # Length regulation
613
+ cond, _, codes, commitment_loss, codebook_loss = inference_module.length_regulator(S_alt, ylens=target_lengths, n_quantizers=3, f0=shifted_f0_alt)
614
+ prompt_condition, _, codes, commitment_loss, codebook_loss = inference_module.length_regulator(S_ori, ylens=target2_lengths, n_quantizers=3, f0=F0_ori)
615
+
616
+ max_source_window = max_context_window - mel2.size(2)
617
+ # split source condition (cond) into chunks
618
+ processed_frames = 0
619
+ generated_wave_chunks = []
620
+ # generate chunk by chunk and stream the output
621
+ while processed_frames < cond.size(1):
622
+ chunk_cond = cond[:, processed_frames:processed_frames + max_source_window]
623
+ is_last_chunk = processed_frames + max_source_window >= cond.size(1)
624
+ cat_condition = torch.cat([prompt_condition, chunk_cond], dim=1)
625
+ with torch.autocast(device_type='cuda', dtype=torch.float16):
626
+ # Voice Conversion
627
+ vc_target = inference_module.cfm.inference(cat_condition,
628
+ torch.LongTensor([cat_condition.size(1)]).to(mel2.device),
629
+ mel2, style2, None, diffusion_steps,
630
+ inference_cfg_rate=inference_cfg_rate)
631
+ vc_target = vc_target[:, :, mel2.size(-1):]
632
+ vc_wave = bigvgan_fn(vc_target.float())[0]
633
+ if processed_frames == 0:
634
+ if is_last_chunk:
635
+ output_wave = vc_wave[0].cpu().numpy()
636
+ generated_wave_chunks.append(output_wave)
637
+ output_wave = (output_wave * 32768.0).astype(np.int16)
638
+ mp3_bytes = AudioSegment(
639
+ output_wave.tobytes(), frame_rate=sr,
640
+ sample_width=output_wave.dtype.itemsize, channels=1
641
+ ).export(format="mp3", bitrate=bitrate).read()
642
+ yield mp3_bytes, (sr, np.concatenate(generated_wave_chunks))
643
+ break
644
+ output_wave = vc_wave[0, :-overlap_wave_len].cpu().numpy()
645
+ generated_wave_chunks.append(output_wave)
646
+ previous_chunk = vc_wave[0, -overlap_wave_len:]
647
+ processed_frames += vc_target.size(2) - overlap_frame_len
648
+ output_wave = (output_wave * 32768.0).astype(np.int16)
649
+ mp3_bytes = AudioSegment(
650
+ output_wave.tobytes(), frame_rate=sr,
651
+ sample_width=output_wave.dtype.itemsize, channels=1
652
+ ).export(format="mp3", bitrate=bitrate).read()
653
+ yield mp3_bytes, None
654
+ elif is_last_chunk:
655
+ output_wave = crossfade(previous_chunk.cpu().numpy(), vc_wave[0].cpu().numpy(), overlap_wave_len)
656
+ generated_wave_chunks.append(output_wave)
657
+ processed_frames += vc_target.size(2) - overlap_frame_len
658
+ output_wave = (output_wave * 32768.0).astype(np.int16)
659
+ mp3_bytes = AudioSegment(
660
+ output_wave.tobytes(), frame_rate=sr,
661
+ sample_width=output_wave.dtype.itemsize, channels=1
662
+ ).export(format="mp3", bitrate=bitrate).read()
663
+ yield mp3_bytes, (sr, np.concatenate(generated_wave_chunks))
664
+ break
665
+ else:
666
+ output_wave = crossfade(previous_chunk.cpu().numpy(), vc_wave[0, :-overlap_wave_len].cpu().numpy(), overlap_wave_len)
667
+ generated_wave_chunks.append(output_wave)
668
+ previous_chunk = vc_wave[0, -overlap_wave_len:]
669
+ processed_frames += vc_target.size(2) - overlap_frame_len
670
+ output_wave = (output_wave * 32768.0).astype(np.int16)
671
+ mp3_bytes = AudioSegment(
672
+ output_wave.tobytes(), frame_rate=sr,
673
+ sample_width=output_wave.dtype.itemsize, channels=1
674
+ ).export(format="mp3", bitrate=bitrate).read()
675
+ yield mp3_bytes, None
676
+
677
+
678
+
679
+
680
+ import gradio as gr
681
+
682
+ gallery_data = {"sikokumetan": {"webp": "default/sikokumetan.webp", "mp3": "default/sikokumetan.mp3"}}
683
+
684
+ def update_reference(evt: gr.SelectData):
685
+ selected_image = evt.value
686
+ for key, value in gallery_data.items():
687
+ if value["webp"] == selected_image:
688
+ print(f"選択された画像: {selected_image}, 対応するMP3: {value['mp3']}")
689
+ return value["mp3"]
690
+ print("対応するMP3が見つかりませんでした。")
691
+ return ""
692
+
693
+ if __name__ == "__main__":
694
+ description = ("Zero-shot音声変換モデル(学習不要)。ローカルでの利用方法は[GitHubリポジトリ](https://github.com/Plachtaa/seed-vc)をご覧ください。"
695
+ "参考音声が25秒を超える場合、自動的に25秒にクリップされます。"
696
+ "また、元音声と参考音声の合計時間が30秒を超える場合、元音声は分割処理されます。")
697
+
698
+ inputs = [
699
+ gr.Audio(type="filepath", label="元音声"),
700
+ gr.Audio(type="filepath", label="参考音声"),
701
+ gr.Slider(minimum=1, maximum=200, value=10, step=1, label="拡散ステップ数", info="デフォルトは10、50~100が最適な品質"),
702
+ gr.Slider(minimum=0.5, maximum=2.0, step=0.1, value=1.0, label="長さ調整", info="1.0未満で速度を上げ、1.0以上で速度を遅くします"),
703
+ gr.Slider(minimum=0.0, maximum=1.0, step=0.1, value=0.7, label="推論CFG率", info="わずかな影響があります"),
704
+ gr.Checkbox(label="F0条件付きモデルを使用", value=False, info="歌声変換には必須です"),
705
+ gr.Checkbox(label="F0自動調整", value=True, info="F0をおおよそ調整して目標音声に合わせます。F0条件付きモデル使用時にのみ有効です"),
706
+ gr.Slider(label='音程変換', minimum=-24, maximum=24, step=1, value=0, info="半音単位の音程変換。F0条件付きモデル使用時にのみ有効です"),
707
+ ]
708
+
709
+ examples = [["examples/source/yae_0.wav", "examples/reference/dingzhen_0.wav", 25, 1.0, 0.7, False, True, 0],
710
+ ["examples/source/jay_0.wav", "examples/reference/azuma_0.wav", 25, 1.0, 0.7, True, True, 0],
711
+ ["examples/source/Wiz Khalifa,Charlie Puth - See You Again [vocals]_[cut_28sec].wav",
712
+ "examples/reference/teio_0.wav", 100, 1.0, 0.7, True, False, 0],
713
+ ["examples/source/TECHNOPOLIS - 2085 [vocals]_[cut_14sec].wav",
714
+ "examples/reference/trump_0.wav", 50, 1.0, 0.7, True, False, -12],
715
+ ]
716
+
717
+ with gr.Blocks() as interface:
718
+ gr.Interface(fn=voice_conversion, description=description, inputs=inputs, outputs=[
719
+ gr.Audio(label="ストリーム出力音声", streaming=True, format='mp3'),
720
+ gr.Audio(label="完全出力音声", streaming=False, format='wav')
721
+ ], title="Seed Voice Conversion", examples=examples, cache_examples=False)
722
+
723
+ with gr.Row():
724
+ gallery = gr.Gallery(label="ギャラリー", show_label=True, value=[gallery_data["sikokumetan"]["webp"]])
725
+
726
+ gallery.select(update_reference, outputs=inputs[1])
727
+
728
+ interface.launch()
729
+ with gr.Blocks() as interface:
730
+ gr.Interface(fn=voice_conversion, description=description, inputs=inputs, outputs=[
731
+ gr.Audio(label="ストリーム出力音声", streaming=True, format='mp3'),
732
+ gr.Audio(label="完全出力音声", streaming=False, format='wav')
733
+ ], title="Seed Voice Conversion", examples=examples, cache_examples=False)
734
+
735
+ with gr.Row():
736
+ gallery = gr.Gallery(label="ギャラリー", show_label=True, value=[gallery_data["sikokumetan"]["webp"]])
737
+
738
+ gallery.select(update_reference, outputs=inputs[1])
739
+
740
+ interface.launch()import gradio as gr
741
 
742
  gallery_data = {"sikokumetan": {"webp": "default/sikokumetan.webp", "mp3": "default/sikokumetan.mp3"}}
743