Spaces:
Sleeping
Sleeping
File size: 7,267 Bytes
46b0a70 |
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 |
from __future__ import annotations
from logging import getLogger
from typing import Any, Literal
import numpy as np
import torch
import torchcrepe
from cm_time import timer
from numpy import dtype, float32, ndarray
from torch import FloatTensor, Tensor
from so_vits_svc_fork.utils import get_optimal_device
LOG = getLogger(__name__)
def normalize_f0(
f0: FloatTensor, x_mask: FloatTensor, uv: FloatTensor, random_scale=True
) -> FloatTensor:
# calculate means based on x_mask
uv_sum = torch.sum(uv, dim=1, keepdim=True)
uv_sum[uv_sum == 0] = 9999
means = torch.sum(f0[:, 0, :] * uv, dim=1, keepdim=True) / uv_sum
if random_scale:
factor = torch.Tensor(f0.shape[0], 1).uniform_(0.8, 1.2).to(f0.device)
else:
factor = torch.ones(f0.shape[0], 1).to(f0.device)
# normalize f0 based on means and factor
f0_norm = (f0 - means.unsqueeze(-1)) * factor.unsqueeze(-1)
if torch.isnan(f0_norm).any():
exit(0)
return f0_norm * x_mask
def interpolate_f0(
f0: ndarray[Any, dtype[float32]]
) -> tuple[ndarray[Any, dtype[float32]], ndarray[Any, dtype[float32]]]:
data = np.reshape(f0, (f0.size, 1))
vuv_vector = np.zeros((data.size, 1), dtype=np.float32)
vuv_vector[data > 0.0] = 1.0
vuv_vector[data <= 0.0] = 0.0
ip_data = data
frame_number = data.size
last_value = 0.0
for i in range(frame_number):
if data[i] <= 0.0:
j = i + 1
for j in range(i + 1, frame_number):
if data[j] > 0.0:
break
if j < frame_number - 1:
if last_value > 0.0:
step = (data[j] - data[i - 1]) / float(j - i)
for k in range(i, j):
ip_data[k] = data[i - 1] + step * (k - i + 1)
else:
for k in range(i, j):
ip_data[k] = data[j]
else:
for k in range(i, frame_number):
ip_data[k] = last_value
else:
ip_data[i] = data[i]
last_value = data[i]
return ip_data[:, 0], vuv_vector[:, 0]
def compute_f0_parselmouth(
wav_numpy: ndarray[Any, dtype[float32]],
p_len: None | int = None,
sampling_rate: int = 44100,
hop_length: int = 512,
):
import parselmouth
x = wav_numpy
if p_len is None:
p_len = x.shape[0] // hop_length
else:
assert abs(p_len - x.shape[0] // hop_length) < 4, "pad length error"
time_step = hop_length / sampling_rate * 1000
f0_min = 50
f0_max = 1100
f0 = (
parselmouth.Sound(x, sampling_rate)
.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")
return f0
def _resize_f0(
x: ndarray[Any, dtype[float32]], target_len: int
) -> ndarray[Any, dtype[float32]]:
source = np.array(x)
source[source < 0.001] = np.nan
target = np.interp(
np.arange(0, len(source) * target_len, len(source)) / target_len,
np.arange(0, len(source)),
source,
)
res = np.nan_to_num(target)
return res
def compute_f0_pyworld(
wav_numpy: ndarray[Any, dtype[float32]],
p_len: None | int = None,
sampling_rate: int = 44100,
hop_length: int = 512,
type_: Literal["dio", "harvest"] = "dio",
):
import pyworld
if p_len is None:
p_len = wav_numpy.shape[0] // hop_length
if type_ == "dio":
f0, t = pyworld.dio(
wav_numpy.astype(np.double),
fs=sampling_rate,
f0_ceil=f0_max,
f0_floor=f0_min,
frame_period=1000 * hop_length / sampling_rate,
)
elif type_ == "harvest":
f0, t = pyworld.harvest(
wav_numpy.astype(np.double),
fs=sampling_rate,
f0_ceil=f0_max,
f0_floor=f0_min,
frame_period=1000 * hop_length / sampling_rate,
)
f0 = pyworld.stonemask(wav_numpy.astype(np.double), f0, t, sampling_rate)
for index, pitch in enumerate(f0):
f0[index] = round(pitch, 1)
return _resize_f0(f0, p_len)
def compute_f0_crepe(
wav_numpy: ndarray[Any, dtype[float32]],
p_len: None | int = None,
sampling_rate: int = 44100,
hop_length: int = 512,
device: str | torch.device = get_optimal_device(),
model: Literal["full", "tiny"] = "full",
):
audio = torch.from_numpy(wav_numpy).to(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()
# (T) -> (1, T)
audio = audio.detach()
pitch: Tensor = torchcrepe.predict(
audio,
sampling_rate,
hop_length,
f0_min,
f0_max,
model,
batch_size=hop_length * 2,
device=device,
pad=True,
)
f0 = pitch.squeeze(0).cpu().float().numpy()
p_len = p_len or wav_numpy.shape[0] // hop_length
f0 = _resize_f0(f0, p_len)
return f0
def compute_f0(
wav_numpy: ndarray[Any, dtype[float32]],
p_len: None | int = None,
sampling_rate: int = 44100,
hop_length: int = 512,
method: Literal["crepe", "crepe-tiny", "parselmouth", "dio", "harvest"] = "dio",
**kwargs,
):
with timer() as t:
wav_numpy = wav_numpy.astype(np.float32)
wav_numpy /= np.quantile(np.abs(wav_numpy), 0.999)
if method in ["dio", "harvest"]:
f0 = compute_f0_pyworld(wav_numpy, p_len, sampling_rate, hop_length, method)
elif method == "crepe":
f0 = compute_f0_crepe(wav_numpy, p_len, sampling_rate, hop_length, **kwargs)
elif method == "crepe-tiny":
f0 = compute_f0_crepe(
wav_numpy, p_len, sampling_rate, hop_length, model="tiny", **kwargs
)
elif method == "parselmouth":
f0 = compute_f0_parselmouth(wav_numpy, p_len, sampling_rate, hop_length)
else:
raise ValueError(
"type must be dio, crepe, crepe-tiny, harvest or parselmouth"
)
rtf = t.elapsed / (len(wav_numpy) / sampling_rate)
LOG.info(f"F0 inference time: {t.elapsed:.3f}s, RTF: {rtf:.3f}")
return f0
def f0_to_coarse(f0: torch.Tensor | float):
is_torch = isinstance(f0, torch.Tensor)
f0_mel = 1127 * (1 + f0 / 700).log() if is_torch else 1127 * np.log(1 + f0 / 700)
f0_mel[f0_mel > 0] = (f0_mel[f0_mel > 0] - f0_mel_min) * (f0_bin - 2) / (
f0_mel_max - f0_mel_min
) + 1
f0_mel[f0_mel <= 1] = 1
f0_mel[f0_mel > f0_bin - 1] = f0_bin - 1
f0_coarse = (f0_mel + 0.5).long() if is_torch else np.rint(f0_mel).astype(np.int)
assert f0_coarse.max() <= 255 and f0_coarse.min() >= 1, (
f0_coarse.max(),
f0_coarse.min(),
)
return f0_coarse
f0_bin = 256
f0_max = 1100.0
f0_min = 50.0
f0_mel_min = 1127 * np.log(1 + f0_min / 700)
f0_mel_max = 1127 * np.log(1 + f0_max / 700)
|