File size: 15,002 Bytes
1cb796d
 
 
 
142eebc
1cb796d
142eebc
1cb796d
 
 
142eebc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1cb796d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
142eebc
 
 
 
 
 
 
 
 
 
 
1cb796d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
142eebc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1cb796d
142eebc
 
 
 
1cb796d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
142eebc
 
1cb796d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
142eebc
1cb796d
 
 
 
142eebc
 
 
1cb796d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
142eebc
 
1cb796d
 
 
 
 
 
 
142eebc
 
 
 
 
 
 
 
1cb796d
 
 
 
142eebc
1cb796d
 
 
 
 
 
142eebc
1cb796d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
142eebc
1cb796d
 
 
 
 
 
 
142eebc
 
 
 
 
 
1cb796d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
142eebc
 
 
 
 
 
 
 
 
1cb796d
 
142eebc
 
1cb796d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
142eebc
 
1cb796d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
142eebc
 
1cb796d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
142eebc
 
1cb796d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
142eebc
 
1cb796d
 
 
142eebc
 
 
 
 
 
 
 
 
 
 
1cb796d
 
 
 
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
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
import numpy as np, parselmouth, torch, pdb
from time import time as ttime
import torch.nn.functional as F
import scipy.signal as signal
import pyworld, os, traceback, faiss, librosa, torchcrepe
from scipy import signal
from functools import lru_cache

bh, ah = signal.butter(N=5, Wn=48, btype="high", fs=16000)

input_audio_path2wav = {}


@lru_cache
def cache_harvest_f0(input_audio_path, fs, f0max, f0min, frame_period):
    audio = input_audio_path2wav[input_audio_path]
    f0, t = pyworld.harvest(
        audio,
        fs=fs,
        f0_ceil=f0max,
        f0_floor=f0min,
        frame_period=frame_period,
    )
    f0 = pyworld.stonemask(audio, f0, t, fs)
    return f0


def change_rms(data1, sr1, data2, sr2, rate):  # 1是输入音频,2是输出音频,rate是2的占比
    # print(data1.max(),data2.max())
    rms1 = librosa.feature.rms(
        y=data1, frame_length=sr1 // 2 * 2, hop_length=sr1 // 2
    )  # 每半秒一个点
    rms2 = librosa.feature.rms(y=data2, frame_length=sr2 // 2 * 2, hop_length=sr2 // 2)
    rms1 = torch.from_numpy(rms1)
    rms1 = F.interpolate(
        rms1.unsqueeze(0), size=data2.shape[0], mode="linear"
    ).squeeze()
    rms2 = torch.from_numpy(rms2)
    rms2 = F.interpolate(
        rms2.unsqueeze(0), size=data2.shape[0], mode="linear"
    ).squeeze()
    rms2 = torch.max(rms2, torch.zeros_like(rms2) + 1e-6)
    data2 *= (
        torch.pow(rms1, torch.tensor(1 - rate))
        * torch.pow(rms2, torch.tensor(rate - 1))
    ).numpy()
    return data2


class VC(object):
    def __init__(self, tgt_sr, config):
        self.x_pad, self.x_query, self.x_center, self.x_max, self.is_half = (
            config.x_pad,
            config.x_query,
            config.x_center,
            config.x_max,
            config.is_half,
        )
        self.sr = 16000  # hubert输入采样率
        self.window = 160  # 每帧点数
        self.t_pad = self.sr * self.x_pad  # 每条前后pad时间
        self.t_pad_tgt = tgt_sr * self.x_pad
        self.t_pad2 = self.t_pad * 2
        self.t_query = self.sr * self.x_query  # 查询切点前后查询时间
        self.t_center = self.sr * self.x_center  # 查询切点位置
        self.t_max = self.sr * self.x_max  # 免查询时长阈值
        self.device = config.device

    def get_f0(
        self,
        input_audio_path,
        x,
        p_len,
        f0_up_key,
        f0_method,
        filter_radius,
        inp_f0=None,
    ):
        global input_audio_path2wav
        time_step = self.window / self.sr * 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)
        if f0_method == "pm":
            f0 = (
                parselmouth.Sound(x, self.sr)
                .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"
                )
        elif f0_method == "harvest":
            input_audio_path2wav[input_audio_path] = x.astype(np.double)
            f0 = cache_harvest_f0(input_audio_path, self.sr, f0_max, f0_min, 10)
            if filter_radius > 2:
                f0 = signal.medfilt(f0, 3)
        elif f0_method == "crepe":
            model = "full"
            # Pick a batch size that doesn't cause memory errors on your gpu
            batch_size = 512
            # Compute pitch using first gpu
            audio = torch.tensor(np.copy(x))[None].float()
            f0, pd = torchcrepe.predict(
                audio,
                self.sr,
                self.window,
                f0_min,
                f0_max,
                model,
                batch_size=batch_size,
                device=self.device,
                return_periodicity=True,
            )
            pd = torchcrepe.filter.median(pd, 3)
            f0 = torchcrepe.filter.mean(f0, 3)
            f0[pd < 0.1] = 0
            f0 = f0[0].cpu().numpy()
        f0 *= pow(2, f0_up_key / 12)
        # with open("test.txt","w")as f:f.write("\n".join([str(i)for i in f0.tolist()]))
        tf0 = self.sr // self.window  # 每秒f0点数
        if inp_f0 is not None:
            delta_t = np.round(
                (inp_f0[:, 0].max() - inp_f0[:, 0].min()) * tf0 + 1
            ).astype("int16")
            replace_f0 = np.interp(
                list(range(delta_t)), inp_f0[:, 0] * 100, inp_f0[:, 1]
            )
            shape = f0[self.x_pad * tf0 : self.x_pad * tf0 + len(replace_f0)].shape[0]
            f0[self.x_pad * tf0 : self.x_pad * tf0 + len(replace_f0)] = replace_f0[
                :shape
            ]
        # with open("test_opt.txt","w")as f:f.write("\n".join([str(i)for i in f0.tolist()]))
        f0bak = f0.copy()
        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, f0bak  # 1-0

    def vc(
        self,
        model,
        net_g,
        sid,
        audio0,
        pitch,
        pitchf,
        times,
        index,
        big_npy,
        index_rate,
        version,
        protect
    ):  # ,file_index,file_big_npy
        feats = torch.from_numpy(audio0)
        if self.is_half:
            feats = feats.half()
        else:
            feats = feats.float()
        if feats.dim() == 2:  # double channels
            feats = feats.mean(-1)
        assert feats.dim() == 1, feats.dim()
        feats = feats.view(1, -1)
        padding_mask = torch.BoolTensor(feats.shape).to(self.device).fill_(False)

        inputs = {
            "source": feats.to(self.device),
            "padding_mask": padding_mask,
            "output_layer": 9 if version == "v1" else 12,
        }
        t0 = ttime()
        with torch.no_grad():
            logits = model.extract_features(**inputs)
            feats = model.final_proj(logits[0]) if version == "v1" else logits[0]
        if(protect<0.5):
            feats0=feats.clone()
        if (
            isinstance(index, type(None)) == False
            and isinstance(big_npy, type(None)) == False
            and index_rate != 0
        ):
            npy = feats[0].cpu().numpy()
            if self.is_half:
                npy = npy.astype("float32")

            # _, I = index.search(npy, 1)
            # npy = big_npy[I.squeeze()]

            score, ix = index.search(npy, k=8)
            weight = np.square(1 / score)
            weight /= weight.sum(axis=1, keepdims=True)
            npy = np.sum(big_npy[ix] * np.expand_dims(weight, axis=2), axis=1)

            if self.is_half:
                npy = npy.astype("float16")
            feats = (
                torch.from_numpy(npy).unsqueeze(0).to(self.device) * index_rate
                + (1 - index_rate) * feats
            )

        feats = F.interpolate(feats.permute(0, 2, 1), scale_factor=2).permute(0, 2, 1)
        if(protect<0.5):
            feats0 = F.interpolate(feats0.permute(0, 2, 1), scale_factor=2).permute(0, 2, 1)
        t1 = ttime()
        p_len = audio0.shape[0] // self.window
        if feats.shape[1] < p_len:
            p_len = feats.shape[1]
            if pitch != None and pitchf != None:
                pitch = pitch[:, :p_len]
                pitchf = pitchf[:, :p_len]

        if(protect<0.5):
            pitchff = pitchf.clone()
            pitchff[pitchf > 0] = 1
            pitchff[pitchf < 1] = protect
            pitchff = pitchff.unsqueeze(-1)
            feats = feats * pitchff + feats0 * (1 - pitchff)
            feats=feats.to(feats0.dtype)
        p_len = torch.tensor([p_len], device=self.device).long()
        with torch.no_grad():
            if pitch != None and pitchf != None:
                audio1 = (
                    (net_g.infer(feats, p_len, pitch, pitchf, sid)[0][0, 0])
                    .data.cpu()
                    .float()
                    .numpy()
                )
            else:
                audio1 = (
                    (net_g.infer(feats, p_len, sid)[0][0, 0]).data.cpu().float().numpy()
                )
        del feats, p_len, padding_mask
        if torch.cuda.is_available():
            torch.cuda.empty_cache()
        t2 = ttime()
        times[0] += t1 - t0
        times[2] += t2 - t1
        return audio1

    def pipeline(
        self,
        model,
        net_g,
        sid,
        audio,
        input_audio_path,
        times,
        f0_up_key,
        f0_method,
        file_index,
        # file_big_npy,
        index_rate,
        if_f0,
        filter_radius,
        tgt_sr,
        resample_sr,
        rms_mix_rate,
        version,
        protect,
        f0_file=None,
    ):
        if (
            file_index != ""
            # and file_big_npy != ""
            # and os.path.exists(file_big_npy) == True
            and os.path.exists(file_index) == True
            and index_rate != 0
        ):
            try:
                index = faiss.read_index(file_index)
                # big_npy = np.load(file_big_npy)
                big_npy = index.reconstruct_n(0, index.ntotal)
            except:
                traceback.print_exc()
                index = big_npy = None
        else:
            index = big_npy = None
        audio = signal.filtfilt(bh, ah, audio)
        audio_pad = np.pad(audio, (self.window // 2, self.window // 2), mode="reflect")
        opt_ts = []
        if audio_pad.shape[0] > self.t_max:
            audio_sum = np.zeros_like(audio)
            for i in range(self.window):
                audio_sum += audio_pad[i : i - self.window]
            for t in range(self.t_center, audio.shape[0], self.t_center):
                opt_ts.append(
                    t
                    - self.t_query
                    + np.where(
                        np.abs(audio_sum[t - self.t_query : t + self.t_query])
                        == np.abs(audio_sum[t - self.t_query : t + self.t_query]).min()
                    )[0][0]
                )
        s = 0
        audio_opt = []
        t = None
        t1 = ttime()
        audio_pad = np.pad(audio, (self.t_pad, self.t_pad), mode="reflect")
        p_len = audio_pad.shape[0] // self.window
        inp_f0 = None
        if hasattr(f0_file, "name") == True:
            try:
                with open(f0_file.name, "r") as f:
                    lines = f.read().strip("\n").split("\n")
                inp_f0 = []
                for line in lines:
                    inp_f0.append([float(i) for i in line.split(",")])
                inp_f0 = np.array(inp_f0, dtype="float32")
            except:
                traceback.print_exc()
        sid = torch.tensor(sid, device=self.device).unsqueeze(0).long()
        pitch, pitchf = None, None
        if if_f0 == 1:
            pitch, pitchf = self.get_f0(
                input_audio_path,
                audio_pad,
                p_len,
                f0_up_key,
                f0_method,
                filter_radius,
                inp_f0,
            )
            pitch = pitch[:p_len]
            pitchf = pitchf[:p_len]
            if self.device == "mps":
                pitchf = pitchf.astype(np.float32)
            pitch = torch.tensor(pitch, device=self.device).unsqueeze(0).long()
            pitchf = torch.tensor(pitchf, device=self.device).unsqueeze(0).float()
        t2 = ttime()
        times[1] += t2 - t1
        for t in opt_ts:
            t = t // self.window * self.window
            if if_f0 == 1:
                audio_opt.append(
                    self.vc(
                        model,
                        net_g,
                        sid,
                        audio_pad[s : t + self.t_pad2 + self.window],
                        pitch[:, s // self.window : (t + self.t_pad2) // self.window],
                        pitchf[:, s // self.window : (t + self.t_pad2) // self.window],
                        times,
                        index,
                        big_npy,
                        index_rate,
                        version,
                        protect
                    )[self.t_pad_tgt : -self.t_pad_tgt]
                )
            else:
                audio_opt.append(
                    self.vc(
                        model,
                        net_g,
                        sid,
                        audio_pad[s : t + self.t_pad2 + self.window],
                        None,
                        None,
                        times,
                        index,
                        big_npy,
                        index_rate,
                        version,
                        protect
                    )[self.t_pad_tgt : -self.t_pad_tgt]
                )
            s = t
        if if_f0 == 1:
            audio_opt.append(
                self.vc(
                    model,
                    net_g,
                    sid,
                    audio_pad[t:],
                    pitch[:, t // self.window :] if t is not None else pitch,
                    pitchf[:, t // self.window :] if t is not None else pitchf,
                    times,
                    index,
                    big_npy,
                    index_rate,
                    version,
                    protect
                )[self.t_pad_tgt : -self.t_pad_tgt]
            )
        else:
            audio_opt.append(
                self.vc(
                    model,
                    net_g,
                    sid,
                    audio_pad[t:],
                    None,
                    None,
                    times,
                    index,
                    big_npy,
                    index_rate,
                    version,
                    protect
                )[self.t_pad_tgt : -self.t_pad_tgt]
            )
        audio_opt = np.concatenate(audio_opt)
        if rms_mix_rate != 1:
            audio_opt = change_rms(audio, 16000, audio_opt, tgt_sr, rms_mix_rate)
        if resample_sr >= 16000 and tgt_sr != resample_sr:
            audio_opt = librosa.resample(
                audio_opt, orig_sr=tgt_sr, target_sr=resample_sr
            )
        audio_max = np.abs(audio_opt).max() / 0.99
        max_int16 = 32768
        if audio_max > 1:
            max_int16 /= audio_max
        audio_opt = (audio_opt * max_int16).astype(np.int16)
        del pitch, pitchf, sid
        if torch.cuda.is_available():
            torch.cuda.empty_cache()
        return audio_opt