File size: 13,886 Bytes
ad16788 |
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 |
from abc import ABC
from abc import abstractmethod
from pathlib import Path
from typing import Collection
from typing import Dict
from typing import Iterable
from typing import Union
import numpy as np
import scipy.signal
import soundfile
from typeguard import check_argument_types
from typeguard import check_return_type
from espnet2.text.build_tokenizer import build_tokenizer
from espnet2.text.cleaner import TextCleaner
from espnet2.text.token_id_converter import TokenIDConverter
class AbsPreprocessor(ABC):
def __init__(self, train: bool):
self.train = train
@abstractmethod
def __call__(
self, uid: str, data: Dict[str, Union[str, np.ndarray]]
) -> Dict[str, np.ndarray]:
raise NotImplementedError
def framing(
x,
frame_length: int = 512,
frame_shift: int = 256,
centered: bool = True,
padded: bool = True,
):
if x.size == 0:
raise ValueError("Input array size is zero")
if frame_length < 1:
raise ValueError("frame_length must be a positive integer")
if frame_length > x.shape[-1]:
raise ValueError("frame_length is greater than input length")
if 0 >= frame_shift:
raise ValueError("frame_shift must be greater than 0")
if centered:
pad_shape = [(0, 0) for _ in range(x.ndim - 1)] + [
(frame_length // 2, frame_length // 2)
]
x = np.pad(x, pad_shape, mode="constant", constant_values=0)
if padded:
# Pad to integer number of windowed segments
# I.e make x.shape[-1] = frame_length + (nseg-1)*nstep,
# with integer nseg
nadd = (-(x.shape[-1] - frame_length) % frame_shift) % frame_length
pad_shape = [(0, 0) for _ in range(x.ndim - 1)] + [(0, nadd)]
x = np.pad(x, pad_shape, mode="constant", constant_values=0)
# Created strided array of data segments
if frame_length == 1 and frame_length == frame_shift:
result = x[..., None]
else:
shape = x.shape[:-1] + (
(x.shape[-1] - frame_length) // frame_shift + 1,
frame_length,
)
strides = x.strides[:-1] + (frame_shift * x.strides[-1], x.strides[-1])
result = np.lib.stride_tricks.as_strided(x, shape=shape, strides=strides)
return result
def detect_non_silence(
x: np.ndarray,
threshold: float = 0.01,
frame_length: int = 1024,
frame_shift: int = 512,
window: str = "boxcar",
) -> np.ndarray:
"""Power based voice activity detection.
Args:
x: (Channel, Time)
>>> x = np.random.randn(1000)
>>> detect = detect_non_silence(x)
>>> assert x.shape == detect.shape
>>> assert detect.dtype == np.bool
"""
if x.shape[-1] < frame_length:
return np.full(x.shape, fill_value=True, dtype=np.bool)
if x.dtype.kind == "i":
x = x.astype(np.float64)
# framed_w: (C, T, F)
framed_w = framing(
x,
frame_length=frame_length,
frame_shift=frame_shift,
centered=False,
padded=True,
)
framed_w *= scipy.signal.get_window(window, frame_length).astype(framed_w.dtype)
# power: (C, T)
power = (framed_w ** 2).mean(axis=-1)
# mean_power: (C,)
mean_power = power.mean(axis=-1)
if np.all(mean_power == 0):
return np.full(x.shape, fill_value=True, dtype=np.bool)
# detect_frames: (C, T)
detect_frames = power / mean_power > threshold
# detects: (C, T, F)
detects = np.broadcast_to(
detect_frames[..., None], detect_frames.shape + (frame_shift,)
)
# detects: (C, TF)
detects = detects.reshape(*detect_frames.shape[:-1], -1)
# detects: (C, TF)
return np.pad(
detects,
[(0, 0)] * (x.ndim - 1) + [(0, x.shape[-1] - detects.shape[-1])],
mode="edge",
)
class CommonPreprocessor(AbsPreprocessor):
def __init__(
self,
train: bool,
token_type: str = None,
token_list: Union[Path, str, Iterable[str]] = None,
bpemodel: Union[Path, str, Iterable[str]] = None,
text_cleaner: Collection[str] = None,
g2p_type: str = None,
unk_symbol: str = "<unk>",
space_symbol: str = "<space>",
non_linguistic_symbols: Union[Path, str, Iterable[str]] = None,
delimiter: str = None,
rir_scp: str = None,
rir_apply_prob: float = 1.0,
noise_scp: str = None,
noise_apply_prob: float = 1.0,
noise_db_range: str = "3_10",
speech_volume_normalize: float = None,
speech_name: str = "speech",
text_name: str = "text",
):
super().__init__(train)
self.train = train
self.speech_name = speech_name
self.text_name = text_name
self.speech_volume_normalize = speech_volume_normalize
self.rir_apply_prob = rir_apply_prob
self.noise_apply_prob = noise_apply_prob
if token_type is not None:
if token_list is None:
raise ValueError("token_list is required if token_type is not None")
self.text_cleaner = TextCleaner(text_cleaner)
self.tokenizer = build_tokenizer(
token_type=token_type,
bpemodel=bpemodel,
delimiter=delimiter,
space_symbol=space_symbol,
non_linguistic_symbols=non_linguistic_symbols,
g2p_type=g2p_type,
)
self.token_id_converter = TokenIDConverter(
token_list=token_list,
unk_symbol=unk_symbol,
)
else:
self.text_cleaner = None
self.tokenizer = None
self.token_id_converter = None
if train and rir_scp is not None:
self.rirs = []
with open(rir_scp, "r", encoding="utf-8") as f:
for line in f:
sps = line.strip().split(None, 1)
if len(sps) == 1:
self.rirs.append(sps[0])
else:
self.rirs.append(sps[1])
else:
self.rirs = None
if train and noise_scp is not None:
self.noises = []
with open(noise_scp, "r", encoding="utf-8") as f:
for line in f:
sps = line.strip().split(None, 1)
if len(sps) == 1:
self.noises.append(sps[0])
else:
self.noises.append(sps[1])
sps = noise_db_range.split("_")
if len(sps) == 1:
self.noise_db_low, self.noise_db_high = float(sps[0])
elif len(sps) == 2:
self.noise_db_low, self.noise_db_high = float(sps[0]), float(sps[1])
else:
raise ValueError(
"Format error: '{noise_db_range}' e.g. -3_4 -> [-3db,4db]"
)
else:
self.noises = None
def __call__(
self, uid: str, data: Dict[str, Union[str, np.ndarray]]
) -> Dict[str, np.ndarray]:
assert check_argument_types()
if self.speech_name in data:
if self.train and self.rirs is not None and self.noises is not None:
speech = data[self.speech_name]
nsamples = len(speech)
# speech: (Nmic, Time)
if speech.ndim == 1:
speech = speech[None, :]
else:
speech = speech.T
# Calc power on non shlence region
power = (speech[detect_non_silence(speech)] ** 2).mean()
# 1. Convolve RIR
if self.rirs is not None and self.rir_apply_prob >= np.random.random():
rir_path = np.random.choice(self.rirs)
if rir_path is not None:
rir, _ = soundfile.read(
rir_path, dtype=np.float64, always_2d=True
)
# rir: (Nmic, Time)
rir = rir.T
# speech: (Nmic, Time)
# Note that this operation doesn't change the signal length
speech = scipy.signal.convolve(speech, rir, mode="full")[
:, : speech.shape[1]
]
# Reverse mean power to the original power
power2 = (speech[detect_non_silence(speech)] ** 2).mean()
speech = np.sqrt(power / max(power2, 1e-10)) * speech
# 2. Add Noise
if (
self.noises is not None
and self.rir_apply_prob >= np.random.random()
):
noise_path = np.random.choice(self.noises)
if noise_path is not None:
noise_db = np.random.uniform(
self.noise_db_low, self.noise_db_high
)
with soundfile.SoundFile(noise_path) as f:
if f.frames == nsamples:
noise = f.read(dtype=np.float64, always_2d=True)
elif f.frames < nsamples:
offset = np.random.randint(0, nsamples - f.frames)
# noise: (Time, Nmic)
noise = f.read(dtype=np.float64, always_2d=True)
# Repeat noise
noise = np.pad(
noise,
[(offset, nsamples - f.frames - offset), (0, 0)],
mode="wrap",
)
else:
offset = np.random.randint(0, f.frames - nsamples)
f.seek(offset)
# noise: (Time, Nmic)
noise = f.read(
nsamples, dtype=np.float64, always_2d=True
)
if len(noise) != nsamples:
raise RuntimeError(f"Something wrong: {noise_path}")
# noise: (Nmic, Time)
noise = noise.T
noise_power = (noise ** 2).mean()
scale = (
10 ** (-noise_db / 20)
* np.sqrt(power)
/ np.sqrt(max(noise_power, 1e-10))
)
speech = speech + scale * noise
speech = speech.T
ma = np.max(np.abs(speech))
if ma > 1.0:
speech /= ma
data[self.speech_name] = speech
if self.speech_volume_normalize is not None:
speech = data[self.speech_name]
ma = np.max(np.abs(speech))
data[self.speech_name] = speech * self.speech_volume_normalize / ma
if self.text_name in data and self.tokenizer is not None:
text = data[self.text_name]
text = self.text_cleaner(text)
tokens = self.tokenizer.text2tokens(text)
text_ints = self.token_id_converter.tokens2ids(tokens)
data[self.text_name] = np.array(text_ints, dtype=np.int64)
assert check_return_type(data)
return data
class CommonPreprocessor_multi(AbsPreprocessor):
def __init__(
self,
train: bool,
token_type: str = None,
token_list: Union[Path, str, Iterable[str]] = None,
bpemodel: Union[Path, str, Iterable[str]] = None,
text_cleaner: Collection[str] = None,
g2p_type: str = None,
unk_symbol: str = "<unk>",
space_symbol: str = "<space>",
non_linguistic_symbols: Union[Path, str, Iterable[str]] = None,
delimiter: str = None,
speech_name: str = "speech",
text_name: list = ["text"],
):
super().__init__(train)
self.train = train
self.speech_name = speech_name
self.text_name = text_name
if token_type is not None:
if token_list is None:
raise ValueError("token_list is required if token_type is not None")
self.text_cleaner = TextCleaner(text_cleaner)
self.tokenizer = build_tokenizer(
token_type=token_type,
bpemodel=bpemodel,
delimiter=delimiter,
space_symbol=space_symbol,
non_linguistic_symbols=non_linguistic_symbols,
g2p_type=g2p_type,
)
self.token_id_converter = TokenIDConverter(
token_list=token_list,
unk_symbol=unk_symbol,
)
else:
self.text_cleaner = None
self.tokenizer = None
self.token_id_converter = None
def __call__(
self, uid: str, data: Dict[str, Union[str, np.ndarray]]
) -> Dict[str, np.ndarray]:
assert check_argument_types()
if self.speech_name in data:
# Nothing now: candidates:
# - STFT
# - Fbank
# - CMVN
# - Data augmentation
pass
for text_n in self.text_name:
if text_n in data and self.tokenizer is not None:
text = data[text_n]
text = self.text_cleaner(text)
tokens = self.tokenizer.text2tokens(text)
text_ints = self.token_id_converter.tokens2ids(tokens)
data[text_n] = np.array(text_ints, dtype=np.int64)
assert check_return_type(data)
return data
|