File size: 9,005 Bytes
bc3753a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
    DeepSpeech features processing routines.
    NB: Based on VOCA code. See the corresponding license restrictions.
"""

__all__ = ['conv_audios_to_deepspeech']

import numpy as np
import warnings
import resampy
from scipy.io import wavfile
from python_speech_features import mfcc
import tensorflow.compat.v1 as tf
tf.disable_v2_behavior()

def conv_audios_to_deepspeech(audios,
                              out_files,
                              num_frames_info,
                              deepspeech_pb_path,
                              audio_window_size=1,
                              audio_window_stride=1):
    """
    Convert list of audio files into files with DeepSpeech features.

    Parameters
    ----------
    audios : list of str or list of None
        Paths to input audio files.
    out_files : list of str
        Paths to output files with DeepSpeech features.
    num_frames_info : list of int
        List of numbers of frames.
    deepspeech_pb_path : str
        Path to DeepSpeech 0.1.0 frozen model.
    audio_window_size : int, default 16
        Audio window size.
    audio_window_stride : int, default 1
        Audio window stride.
    """
    # deepspeech_pb_path="/disk4/keyu/DeepSpeech/deepspeech-0.9.2-models.pbmm"
    graph, logits_ph, input_node_ph, input_lengths_ph = prepare_deepspeech_net(
        deepspeech_pb_path)

    with tf.compat.v1.Session(graph=graph) as sess:
        for audio_file_path, out_file_path, num_frames in zip(audios, out_files, num_frames_info):
            print(audio_file_path)
            print(out_file_path)
            audio_sample_rate, audio = wavfile.read(audio_file_path)
            if audio.ndim != 1:
                warnings.warn(
                    "Audio has multiple channels, the first channel is used")
                audio = audio[:, 0]
            ds_features = pure_conv_audio_to_deepspeech(
                audio=audio,
                audio_sample_rate=audio_sample_rate,
                audio_window_size=audio_window_size,
                audio_window_stride=audio_window_stride,
                num_frames=num_frames,
                net_fn=lambda x: sess.run(
                    logits_ph,
                    feed_dict={
                        input_node_ph: x[np.newaxis, ...],
                        input_lengths_ph: [x.shape[0]]}))

            net_output = ds_features.reshape(-1, 29)
            win_size = 16
            zero_pad = np.zeros((int(win_size / 2), net_output.shape[1]))
            net_output = np.concatenate(
                (zero_pad, net_output, zero_pad), axis=0)
            windows = []
            for window_index in range(0, net_output.shape[0] - win_size, 2):
                windows.append(
                    net_output[window_index:window_index + win_size])
            print(np.array(windows).shape)
            np.save(out_file_path, np.array(windows))


def prepare_deepspeech_net(deepspeech_pb_path):
    """
    Load and prepare DeepSpeech network.

    Parameters
    ----------
    deepspeech_pb_path : str
        Path to DeepSpeech 0.1.0 frozen model.

    Returns
    -------
    graph : obj
        ThensorFlow graph.
    logits_ph : obj
        ThensorFlow placeholder for `logits`.
    input_node_ph : obj
        ThensorFlow placeholder for `input_node`.
    input_lengths_ph : obj
        ThensorFlow placeholder for `input_lengths`.
    """
    # Load graph and place_holders:
    with tf.io.gfile.GFile(deepspeech_pb_path, "rb") as f:
        graph_def = tf.compat.v1.GraphDef()
        graph_def.ParseFromString(f.read())

    graph = tf.compat.v1.get_default_graph()
    tf.import_graph_def(graph_def, name="deepspeech")
    logits_ph = graph.get_tensor_by_name("deepspeech/logits:0")
    input_node_ph = graph.get_tensor_by_name("deepspeech/input_node:0")
    input_lengths_ph = graph.get_tensor_by_name("deepspeech/input_lengths:0")

    return graph, logits_ph, input_node_ph, input_lengths_ph


def pure_conv_audio_to_deepspeech(audio,
                                  audio_sample_rate,
                                  audio_window_size,
                                  audio_window_stride,
                                  num_frames,
                                  net_fn):
    """
    Core routine for converting audion into DeepSpeech features.

    Parameters
    ----------
    audio : np.array
        Audio data.
    audio_sample_rate : int
        Audio sample rate.
    audio_window_size : int
        Audio window size.
    audio_window_stride : int
        Audio window stride.
    num_frames : int or None
        Numbers of frames.
    net_fn : func
        Function for DeepSpeech model call.

    Returns
    -------
    np.array
        DeepSpeech features.
    """
    target_sample_rate = 16000
    if audio_sample_rate != target_sample_rate:
        resampled_audio = resampy.resample(
            x=audio.astype(np.float),
            sr_orig=audio_sample_rate,
            sr_new=target_sample_rate)
    else:
        resampled_audio = audio.astype(np.float32)
    input_vector = conv_audio_to_deepspeech_input_vector(
        audio=resampled_audio.astype(np.int16),
        sample_rate=target_sample_rate,
        num_cepstrum=26,
        num_context=9)

    network_output = net_fn(input_vector)
    # print(network_output.shape)

    deepspeech_fps = 50
    video_fps = 50  # Change this option if video fps is different
    audio_len_s = float(audio.shape[0]) / audio_sample_rate
    if num_frames is None:
        num_frames = int(round(audio_len_s * video_fps))
    else:
        video_fps = num_frames / audio_len_s
    network_output = interpolate_features(
        features=network_output[:, 0],
        input_rate=deepspeech_fps,
        output_rate=video_fps,
        output_len=num_frames)

    # Make windows:
    zero_pad = np.zeros((int(audio_window_size / 2), network_output.shape[1]))
    network_output = np.concatenate(
        (zero_pad, network_output, zero_pad), axis=0)
    windows = []
    for window_index in range(0, network_output.shape[0] - audio_window_size, audio_window_stride):
        windows.append(
            network_output[window_index:window_index + audio_window_size])

    return np.array(windows)


def conv_audio_to_deepspeech_input_vector(audio,
                                          sample_rate,
                                          num_cepstrum,
                                          num_context):
    """
    Convert audio raw data into DeepSpeech input vector.

    Parameters
    ----------
    audio : np.array
        Audio data.
    audio_sample_rate : int
        Audio sample rate.
    num_cepstrum : int
        Number of cepstrum.
    num_context : int
        Number of context.

    Returns
    -------
    np.array
        DeepSpeech input vector.
    """
    # Get mfcc coefficients:
    features = mfcc(
        signal=audio,
        samplerate=sample_rate,
        numcep=num_cepstrum)

    # We only keep every second feature (BiRNN stride = 2):
    features = features[::2]

    # One stride per time step in the input:
    num_strides = len(features)

    # Add empty initial and final contexts:
    empty_context = np.zeros((num_context, num_cepstrum), dtype=features.dtype)
    features = np.concatenate((empty_context, features, empty_context))

    # Create a view into the array with overlapping strides of size
    # numcontext (past) + 1 (present) + numcontext (future):
    window_size = 2 * num_context + 1
    train_inputs = np.lib.stride_tricks.as_strided(
        features,
        shape=(num_strides, window_size, num_cepstrum),
        strides=(features.strides[0],
                 features.strides[0], features.strides[1]),
        writeable=False)

    # Flatten the second and third dimensions:
    train_inputs = np.reshape(train_inputs, [num_strides, -1])

    train_inputs = np.copy(train_inputs)
    train_inputs = (train_inputs - np.mean(train_inputs)) / \
        np.std(train_inputs)

    return train_inputs


def interpolate_features(features,
                         input_rate,
                         output_rate,
                         output_len):
    """
    Interpolate DeepSpeech features.

    Parameters
    ----------
    features : np.array
        DeepSpeech features.
    input_rate : int
        input rate (FPS).
    output_rate : int
        Output rate (FPS).
    output_len : int
        Output data length.

    Returns
    -------
    np.array
        Interpolated data.
    """
    input_len = features.shape[0]
    num_features = features.shape[1]
    input_timestamps = np.arange(input_len) / float(input_rate)
    output_timestamps = np.arange(output_len) / float(output_rate)
    output_features = np.zeros((output_len, num_features))
    for feature_idx in range(num_features):
        output_features[:, feature_idx] = np.interp(
            x=output_timestamps,
            xp=input_timestamps,
            fp=features[:, feature_idx])
    return output_features