Spaces:
Running
Running
File size: 21,828 Bytes
55adc26 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 |
import numpy as np, parselmouth, torch, pdb, sys, os
from time import time as ttime
import torch.nn.functional as F
import torchcrepe
from torch import Tensor
import scipy.signal as signal
import pyworld, os, faiss, librosa, torchcrepe
from scipy import signal
from functools import lru_cache
import random
import gc
import re
now_dir = os.getcwd()
sys.path.append(now_dir)
from rvc.lib.FCPEF0Predictor import FCPEF0Predictor
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):
# print(data1.max(),data2.max())
rms1 = librosa.feature.rms(y=data1, frame_length=sr1 // 2 * 2, hop_length=sr1 // 2)
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
self.window = 160
self.t_pad = self.sr * self.x_pad
self.t_pad_tgt = tgt_sr * self.x_pad
self.t_pad2 = self.t_pad * 2
self.t_query = self.sr * self.x_query
self.t_center = self.sr * self.x_center
self.t_max = self.sr * self.x_max
self.device = config.device
self.ref_freqs = [
65.41,
82.41,
110.00,
146.83,
196.00,
246.94,
329.63,
440.00,
587.33,
783.99,
1046.50,
]
# Generate interpolated frequencies
self.note_dict = self.generate_interpolated_frequencies()
def generate_interpolated_frequencies(self):
# Generate interpolated frequencies based on the reference frequencies.
note_dict = []
for i in range(len(self.ref_freqs) - 1):
freq_low = self.ref_freqs[i]
freq_high = self.ref_freqs[i + 1]
# Interpolate between adjacent reference frequencies
interpolated_freqs = np.linspace(
freq_low, freq_high, num=10, endpoint=False
)
note_dict.extend(interpolated_freqs)
# Add the last reference frequency
note_dict.append(self.ref_freqs[-1])
return note_dict
def autotune_f0(self, f0):
# Autotunes the given fundamental frequency (f0) to the nearest musical note.
autotuned_f0 = np.zeros_like(f0)
for i, freq in enumerate(f0):
# Find the closest note
closest_note = min(self.note_dict, key=lambda x: abs(x - freq))
autotuned_f0[i] = closest_note
return autotuned_f0
def get_optimal_torch_device(self, index: int = 0) -> torch.device:
if torch.cuda.is_available():
return torch.device(f"cuda:{index % torch.cuda.device_count()}")
elif torch.backends.mps.is_available():
return torch.device("mps")
return torch.device("cpu")
def get_f0_crepe_computation(
self,
x,
f0_min,
f0_max,
p_len,
hop_length,
model="full",
):
x = x.astype(np.float32)
x /= np.quantile(np.abs(x), 0.999)
torch_device = self.get_optimal_torch_device()
audio = torch.from_numpy(x).to(torch_device, copy=True)
audio = torch.unsqueeze(audio, dim=0)
if audio.ndim == 2 and audio.shape[0] > 1:
audio = torch.mean(audio, dim=0, keepdim=True).detach()
audio = audio.detach()
pitch: Tensor = torchcrepe.predict(
audio,
self.sr,
hop_length,
f0_min,
f0_max,
model,
batch_size=hop_length * 2,
device=torch_device,
pad=True,
)
p_len = p_len or x.shape[0] // hop_length
source = np.array(pitch.squeeze(0).cpu().float().numpy())
source[source < 0.001] = np.nan
target = np.interp(
np.arange(0, len(source) * p_len, len(source)) / p_len,
np.arange(0, len(source)),
source,
)
f0 = np.nan_to_num(target)
return f0
def get_f0_official_crepe_computation(
self,
x,
f0_min,
f0_max,
model="full",
):
batch_size = 512
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()
return f0
def get_f0_hybrid_computation(
self,
methods_str,
x,
f0_min,
f0_max,
p_len,
hop_length,
):
methods_str = re.search("hybrid\[(.+)\]", methods_str)
if methods_str:
methods = [method.strip() for method in methods_str.group(1).split("+")]
f0_computation_stack = []
print(f"Calculating f0 pitch estimations for methods {str(methods)}")
x = x.astype(np.float32)
x /= np.quantile(np.abs(x), 0.999)
for method in methods:
f0 = None
if method == "crepe":
f0 = self.get_f0_crepe_computation(
x, f0_min, f0_max, p_len, int(hop_length)
)
elif method == "rmvpe":
if hasattr(self, "model_rmvpe") == False:
from rvc.lib.rmvpe import RMVPE
self.model_rmvpe = RMVPE(
"rmvpe.pt", is_half=self.is_half, device=self.device
)
f0 = self.model_rmvpe.infer_from_audio(x, thred=0.03)
f0 = f0[1:]
elif method == "fcpe":
self.model_fcpe = FCPEF0Predictor(
"fcpe.pt",
f0_min=int(f0_min),
f0_max=int(f0_max),
dtype=torch.float32,
device=self.device,
sampling_rate=self.sr,
threshold=0.03,
)
f0 = self.model_fcpe.compute_f0(x, p_len=p_len)
del self.model_fcpe
gc.collect()
f0_computation_stack.append(f0)
print(f"Calculating hybrid median f0 from the stack of {str(methods)}")
f0_computation_stack = [fc for fc in f0_computation_stack if fc is not None]
f0_median_hybrid = None
if len(f0_computation_stack) == 1:
f0_median_hybrid = f0_computation_stack[0]
else:
f0_median_hybrid = np.nanmedian(f0_computation_stack, axis=0)
return f0_median_hybrid
def get_f0(
self,
input_audio_path,
x,
p_len,
f0_up_key,
f0_method,
filter_radius,
hop_length,
f0autotune,
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 int(filter_radius) > 2:
f0 = signal.medfilt(f0, 3)
elif f0_method == "dio":
f0, t = pyworld.dio(
x.astype(np.double),
fs=self.sr,
f0_ceil=f0_max,
f0_floor=f0_min,
frame_period=10,
)
f0 = pyworld.stonemask(x.astype(np.double), f0, t, self.sr)
f0 = signal.medfilt(f0, 3)
elif f0_method == "crepe":
f0 = self.get_f0_crepe_computation(
x, f0_min, f0_max, p_len, int(hop_length)
)
elif f0_method == "crepe-tiny":
f0 = self.get_f0_crepe_computation(
x, f0_min, f0_max, p_len, int(hop_length), "tiny"
)
elif f0_method == "rmvpe":
if hasattr(self, "model_rmvpe") == False:
from rvc.lib.rmvpe import RMVPE
self.model_rmvpe = RMVPE(
"rmvpe.pt", is_half=self.is_half, device=self.device
)
f0 = self.model_rmvpe.infer_from_audio(x, thred=0.03)
elif f0_method == "fcpe":
self.model_fcpe = FCPEF0Predictor(
"fcpe.pt",
f0_min=int(f0_min),
f0_max=int(f0_max),
dtype=torch.float32,
device=self.device,
sampling_rate=self.sr,
threshold=0.03,
)
f0 = self.model_fcpe.compute_f0(x, p_len=p_len)
del self.model_fcpe
gc.collect()
elif "hybrid" in f0_method:
input_audio_path2wav[input_audio_path] = x.astype(np.double)
f0 = self.get_f0_hybrid_computation(
f0_method,
x,
f0_min,
f0_max,
p_len,
hop_length,
)
if f0autotune == "True":
f0 = self.autotune_f0(f0)
f0 *= pow(2, f0_up_key / 12)
tf0 = self.sr // self.window
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
]
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
def vc(
self,
model,
net_g,
sid,
audio0,
pitch,
pitchf,
index,
big_npy,
index_rate,
version,
protect,
):
feats = torch.from_numpy(audio0)
if self.is_half:
feats = feats.half()
else:
feats = feats.float()
if feats.dim() == 2:
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 and pitch != None and pitchf != None:
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")
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 and pitch != None and pitchf != None:
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 and pitch != None and pitchf != None:
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()
return audio1
def pipeline(
self,
model,
net_g,
sid,
audio,
input_audio_path,
f0_up_key,
f0_method,
file_index,
index_rate,
if_f0,
filter_radius,
tgt_sr,
resample_sr,
rms_mix_rate,
version,
protect,
hop_length,
f0autotune,
f0_file=None,
):
if file_index != "" and os.path.exists(file_index) == True and index_rate != 0:
try:
index = faiss.read_index(file_index)
big_npy = index.reconstruct_n(0, index.ntotal)
except Exception as error:
print(error)
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 Exception as error:
print(error)
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,
hop_length,
f0autotune,
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()
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],
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,
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,
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,
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
|