forked from huggingface/transformers
-
Notifications
You must be signed in to change notification settings - Fork 0
/
feature_extraction_whisper.py
320 lines (257 loc) 路 12.9 KB
/
feature_extraction_whisper.py
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
# coding=utf-8
# Copyright 2022 The HuggingFace Inc. team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Feature extractor class for Whisper
"""
from typing import List, Optional, Union
import numpy as np
from numpy.fft import fft
from ...feature_extraction_sequence_utils import SequenceFeatureExtractor
from ...feature_extraction_utils import BatchFeature
from ...utils import TensorType, logging
logger = logging.get_logger(__name__)
class WhisperFeatureExtractor(SequenceFeatureExtractor):
r"""
Constructs a Whisper feature extractor.
This feature extractor inherits from [`WhisperFeatureExtractor`] which contains most of the main methods. Users
should refer to this superclass for more information regarding those methods.
This class extracts mel-filter bank features from raw speech using a custom numpy implementation of the `Short Time
Fourier Transform` which should match pytorch's `torch.stft` equivalent.
Args:
feature_size (`int`, defaults to 80):
The feature dimension of the extracted features.
sampling_rate (`int`, defaults to 16000):
The sampling rate at which the audio files should be digitalized expressed in Hertz per second (Hz).
hop_length (`int`, defaults to 160):
Length of the overlaping windows for the STFT used to obtain the Mel Frequency coefficients.
chunk_length (`int`, defaults to 30):
The maximum number of chuncks of `sampling_rate` samples used to trim and pad longer or shorter audio
sequences.
n_fft (`int`, defaults to 400):
Size of the Fourier transform.
padding_value (`float`, *optional*, defaults to 0.0):
Padding value used to pad the audio. Should correspond to silences.
"""
model_input_names = ["input_features"]
def __init__(
self,
feature_size=80,
sampling_rate=16000,
hop_length=160,
chunk_length=30,
n_fft=400,
padding_value=0.0,
**kwargs
):
super().__init__(feature_size=feature_size, sampling_rate=sampling_rate, padding_value=padding_value, **kwargs)
self.n_fft = n_fft
self.hop_length = hop_length
self.chunk_length = chunk_length
self.return_attention_mask = True
self.n_samples = chunk_length * sampling_rate
self.nb_max_frames = self.n_samples // hop_length
self.sampling_rate = sampling_rate
self.mel_filters = self.get_mel_filters(sampling_rate, n_fft, n_mels=feature_size)
def get_mel_filters(self, sr, n_fft, n_mels=128, dtype=np.float32):
# Initialize the weights
n_mels = int(n_mels)
weights = np.zeros((n_mels, int(1 + n_fft // 2)), dtype=dtype)
# Center freqs of each FFT bin
fftfreqs = np.fft.rfftfreq(n=n_fft, d=1.0 / sr)
# 'Center freqs' of mel bands - uniformly spaced between limits
min_mel = 0.0
max_mel = 45.245640471924965
mels = np.linspace(min_mel, max_mel, n_mels + 2)
mels = np.asanyarray(mels)
# Fill in the linear scale
f_min = 0.0
f_sp = 200.0 / 3
freqs = f_min + f_sp * mels
# And now the nonlinear scale
min_log_hz = 1000.0 # beginning of log region (Hz)
min_log_mel = (min_log_hz - f_min) / f_sp # same (Mels)
logstep = np.log(6.4) / 27.0 # step size for log region
# If we have vector data, vectorize
log_t = mels >= min_log_mel
freqs[log_t] = min_log_hz * np.exp(logstep * (mels[log_t] - min_log_mel))
mel_f = freqs
fdiff = np.diff(mel_f)
ramps = np.subtract.outer(mel_f, fftfreqs)
for i in range(n_mels):
# lower and upper slopes for all bins
lower = -ramps[i] / fdiff[i]
upper = ramps[i + 2] / fdiff[i + 1]
# .. then intersect them with each other and zero
weights[i] = np.maximum(0, np.minimum(lower, upper))
# Slaney-style mel is scaled to be approx constant energy per channel
enorm = 2.0 / (mel_f[2 : n_mels + 2] - mel_f[:n_mels])
weights *= enorm[:, np.newaxis]
return weights
def fram_wave(self, waveform, center=True):
"""
Transform a raw waveform into a list of smaller waveforms. The window length defines how much of the signal is
contain in each frame (smalle waveform), while the hope length defines the step between the beginning of each
new frame.
Centering is done by reflecting the waveform which is first centered around `frame_idx * hop_length`.
"""
frames = []
for i in range(0, waveform.shape[0] + 1, self.hop_length):
half_window = (self.n_fft - 1) // 2 + 1
if center:
start = i - half_window if i > half_window else 0
end = i + half_window if i < waveform.shape[0] - half_window else waveform.shape[0]
frame = waveform[start:end]
if start == 0:
padd_width = (-i + half_window, 0)
frame = np.pad(frame, pad_width=padd_width, mode="reflect")
elif end == waveform.shape[0]:
padd_width = (0, (i - waveform.shape[0] + half_window))
frame = np.pad(frame, pad_width=padd_width, mode="reflect")
else:
frame = waveform[i : i + self.n_fft]
frame_width = frame.shape[0]
if frame_width < waveform.shape[0]:
frame = np.lib.pad(
frame, pad_width=(0, self.n_fft - frame_width), mode="constant", constant_values=0
)
frames.append(frame)
return np.stack(frames, 0)
def stft(self, frames, window):
"""
Calculates the complex Short-Time Fourier Transform (STFT) of the given framed signal. Should give the same
results as `torch.stft`.
"""
frame_size = frames.shape[1]
fft_size = self.n_fft
if fft_size is None:
fft_size = frame_size
if fft_size < frame_size:
raise ValueError("FFT size must greater or equal the frame size")
# number of FFT bins to store
num_fft_bins = (fft_size >> 1) + 1
data = np.empty((len(frames), num_fft_bins), dtype=np.complex64)
fft_signal = np.zeros(fft_size)
for f, frame in enumerate(frames):
if window is not None:
np.multiply(frame, window, out=fft_signal[:frame_size])
else:
fft_signal[:frame_size] = frame
data[f] = fft(fft_signal, axis=0)[:num_fft_bins]
return data.T
def _np_extract_fbank_features(self, waveform: np.array) -> np.ndarray:
"""
Compute the log-Mel spectrogram of the provided audio, gives similar results whisper's original torch
implementation with 1e-5 tolerance.
"""
window = np.hanning(self.n_fft + 1)[:-1]
frames = self.fram_wave(waveform)
stft = self.stft(frames, window=window)
magnitudes = np.abs(stft[:, :-1]) ** 2
filters = self.mel_filters
mel_spec = filters @ magnitudes
log_spec = np.log10(np.clip(mel_spec, a_min=1e-10, a_max=None))
log_spec = np.maximum(log_spec, log_spec.max() - 8.0)
log_spec = (log_spec + 4.0) / 4.0
return log_spec
def __call__(
self,
raw_speech: Union[np.ndarray, List[float], List[np.ndarray], List[List[float]]],
truncation: bool = True,
pad_to_multiple_of: Optional[int] = None,
return_tensors: Optional[Union[str, TensorType]] = None,
return_attention_mask: Optional[bool] = None,
padding: Optional[str] = "max_length",
max_length: Optional[int] = None,
sampling_rate: Optional[int] = None,
**kwargs
) -> BatchFeature:
"""
Main method to featurize and prepare for the model one or several sequence(s).
Args:
raw_speech (`np.ndarray`, `List[float]`, `List[np.ndarray]`, `List[List[float]]`):
The sequence or batch of sequences to be padded. Each sequence can be a numpy array, a list of float
values, a list of numpy arrays or a list of list of float values.
truncation (`bool`, *optional*, default to `True`):
Activates truncation to cut input sequences longer than *max_length* to *max_length*.
pad_to_multiple_of (`int`, *optional*, defaults to None):
If set will pad the sequence to a multiple of the provided value.
This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability
>= 7.5 (Volta), or on TPUs which benefit from having sequence lengths be a multiple of 128.
return_attention_mask (`bool`, *optional*):
Whether to return the attention mask. If left to the default, will return the attention mask according
to the specific feature_extractor's default.
[What are attention masks?](../glossary#attention-mask)
<Tip>
For WhisperTransoformer models, `attention_mask` should alwys be passed for batched inference, to avoid
subtle bugs.
</Tip>
return_tensors (`str` or [`~utils.TensorType`], *optional*):
If set, will return tensors instead of list of python integers. Acceptable values are:
- `'tf'`: Return TensorFlow `tf.constant` objects.
- `'pt'`: Return PyTorch `torch.Tensor` objects.
- `'np'`: Return Numpy `np.ndarray` objects.
sampling_rate (`int`, *optional*):
The sampling rate at which the `raw_speech` input was sampled. It is strongly recommended to pass
`sampling_rate` at the forward call to prevent silent errors and allow automatic speech recognition
pipeline.
padding_value (`float`, defaults to 0.0):
The value that is used to fill the padding values / vectors.
"""
if sampling_rate is not None:
if sampling_rate != self.sampling_rate:
raise ValueError(
f"The model corresponding to this feature extractor: {self} was trained using a sampling rate of"
f" {self.sampling_rate}. Please make sure that the provided `raw_speech` input was sampled with"
f" {self.sampling_rate} and not {sampling_rate}."
)
else:
logger.warning(
"It is strongly recommended to pass the `sampling_rate` argument to this function. "
"Failing to do so can result in silent errors that might be hard to debug."
)
is_batched = bool(
isinstance(raw_speech, (list, tuple))
and (isinstance(raw_speech[0], np.ndarray) or isinstance(raw_speech[0], (tuple, list)))
)
if is_batched:
raw_speech = [np.asarray([speech], dtype=np.float32).T for speech in raw_speech]
elif not is_batched and not isinstance(raw_speech, np.ndarray):
raw_speech = np.asarray(raw_speech, dtype=np.float32)
elif isinstance(raw_speech, np.ndarray) and raw_speech.dtype is np.dtype(np.float64):
raw_speech = raw_speech.astype(np.float32)
# always return batch
if not is_batched:
raw_speech = [np.asarray([raw_speech]).T]
batched_speech = BatchFeature({"input_features": raw_speech})
# convert into correct format for padding
padded_inputs = self.pad(
batched_speech,
padding=padding,
max_length=max_length if max_length else self.n_samples,
truncation=truncation,
pad_to_multiple_of=pad_to_multiple_of,
return_attention_mask=False,
**kwargs,
)
# make sure list is in array format
input_features = padded_inputs.get("input_features").transpose(2, 0, 1)
input_features = [self._np_extract_fbank_features(waveform) for waveform in input_features[0]]
if isinstance(input_features[0], List):
padded_inputs["input_features"] = [np.asarray(feature, dtype=np.float32) for feature in input_features]
else:
padded_inputs["input_features"] = input_features
if return_tensors is not None:
padded_inputs = padded_inputs.convert_to_tensors(return_tensors)
return padded_inputs