mirror of
https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI.git
synced 2026-08-29 10:09:32 +02:00
387 lines
13 KiB
Python
387 lines
13 KiB
Python
import os
|
||
import traceback
|
||
import logging
|
||
|
||
logger = logging.getLogger(__name__)
|
||
|
||
from time import time as ttime
|
||
|
||
import faiss
|
||
import librosa
|
||
import numpy as np
|
||
import parselmouth
|
||
import torch
|
||
import torch.nn.functional as F
|
||
from scipy import signal
|
||
|
||
from infer.hubert import extract_hubert_features
|
||
|
||
bh, ah = signal.butter(N=5, Wn=48, btype="high", fs=16000)
|
||
|
||
|
||
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 Pipeline(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,
|
||
x,
|
||
p_len,
|
||
f0_up_key,
|
||
f0_method,
|
||
):
|
||
if f0_method not in ("pm", "rmvpe", "fcpe"):
|
||
raise ValueError(f"Unsupported F0 method: {f0_method}")
|
||
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 == "rmvpe":
|
||
if not hasattr(self, "model_rmvpe"):
|
||
from infer.rmvpe import RMVPE
|
||
|
||
logger.info(
|
||
"Loading rmvpe model,%s" % "%s/rmvpe.pt" % os.environ["rmvpe_root"]
|
||
)
|
||
self.model_rmvpe = RMVPE(
|
||
"%s/rmvpe.pt" % os.environ["rmvpe_root"],
|
||
is_half=self.is_half,
|
||
device=self.device,
|
||
)
|
||
f0 = self.model_rmvpe.infer_from_audio(x, thred=0.03)
|
||
|
||
if "privateuseone" in str(self.device): # clean ortruntime memory
|
||
del self.model_rmvpe.model
|
||
del self.model_rmvpe
|
||
logger.info("Cleaning ortruntime memory")
|
||
elif f0_method == "fcpe":
|
||
if not hasattr(self, "model_fcpe"):
|
||
from infer.fcpe import FCPEInfer
|
||
|
||
logger.info("Loading fcpe model")
|
||
self.model_fcpe = FCPEInfer(self.device)
|
||
f0 = self.model_fcpe.infer(
|
||
torch.from_numpy(x).unsqueeze(0).float(),
|
||
sr=self.sr,
|
||
decoder_mode="local_argmax",
|
||
threshold=0.006,
|
||
).squeeze().detach().cpu().numpy()
|
||
|
||
try:
|
||
uv = f0 == 0
|
||
f0[uv] = np.interp(np.where(uv)[0], np.where(~uv)[0], f0[~uv])
|
||
except Exception:
|
||
traceback.print_exc()
|
||
f0 *= pow(2, f0_up_key / 12)
|
||
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.int32)
|
||
return f0_coarse, f0bak # 1-0
|
||
|
||
def vc(
|
||
self,
|
||
model,
|
||
net_g,
|
||
sid,
|
||
audio0,
|
||
pitch,
|
||
pitchf,
|
||
times,
|
||
index,
|
||
index_vectors,
|
||
index_rate,
|
||
version,
|
||
protect,
|
||
):
|
||
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)
|
||
|
||
t0 = ttime()
|
||
with torch.no_grad():
|
||
feats = extract_hubert_features(
|
||
model,
|
||
feats.to(self.device),
|
||
version,
|
||
padding_mask=padding_mask,
|
||
)
|
||
if protect < 0.5 and pitch is not None and pitchf is not None:
|
||
feats0 = feats.clone()
|
||
if (
|
||
not isinstance(index, type(None))
|
||
and not isinstance(index_vectors, type(None))
|
||
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(index_vectors[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 is not None and pitchf is not 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 is not None and pitchf is not None:
|
||
pitch = pitch[:, :p_len]
|
||
pitchf = pitchf[:, :p_len]
|
||
|
||
if protect < 0.5 and pitch is not None and pitchf is not 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():
|
||
hasp = pitch is not None and pitchf is not None
|
||
arg = (feats, p_len, pitch, pitchf, sid) if hasp else (feats, p_len, sid)
|
||
audio1 = (net_g.infer(*arg)[0][0, 0]).data.cpu().float().numpy()
|
||
del hasp, arg
|
||
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,
|
||
times,
|
||
f0_up_key,
|
||
f0_method,
|
||
file_index,
|
||
index_rate,
|
||
if_f0,
|
||
tgt_sr,
|
||
resample_sr,
|
||
rms_mix_rate,
|
||
version,
|
||
protect,
|
||
):
|
||
if (
|
||
file_index != ""
|
||
and os.path.exists(file_index)
|
||
and index_rate != 0
|
||
):
|
||
try:
|
||
index = faiss.read_index(file_index)
|
||
index_vectors = index.reconstruct_n(0, index.ntotal)
|
||
except:
|
||
traceback.print_exc()
|
||
index = index_vectors = None
|
||
else:
|
||
index = index_vectors = 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 += np.abs(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(
|
||
audio_sum[t - self.t_query : t + self.t_query]
|
||
== 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
|
||
sid = torch.tensor(sid, device=self.device).unsqueeze(0).long()
|
||
pitch, pitchf = None, None
|
||
if if_f0 == 1:
|
||
pitch, pitchf = self.get_f0(
|
||
audio_pad,
|
||
p_len,
|
||
f0_up_key,
|
||
f0_method,
|
||
)
|
||
pitch = pitch[:p_len]
|
||
pitchf = pitchf[:p_len]
|
||
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,
|
||
index_vectors,
|
||
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,
|
||
index_vectors,
|
||
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,
|
||
index_vectors,
|
||
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,
|
||
index_vectors,
|
||
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 tgt_sr != resample_sr >= 16000:
|
||
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
|