onnx-asr/preprocessors/nemo.py at ffc1ed0b170e658eba6f5b1aa8960ef7eb38e4dc · istupakov/onnx-asr · GitHub

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
"""LogMelSpectrogram feature extractor for Nemo models."""

import numpy as np
from onnxscript import FLOAT, INT64, script
from onnxscript import opset17 as op

from preprocessors.fbanks import melscale_fbanks
from preprocessors.stft import conv_power_spectrogram, stft_conv_weights

sample_rate = 16_000
n_fft = 512
win_length = 400
hop_length = 160
preemph = 0.97

log_zero_guard_value = float(2**-24)

melscale_fbanks80 = melscale_fbanks(n_fft // 2 + 1, 0, sample_rate // 2, 80, sample_rate, "slaney", "slaney").astype(
    np.float32
)
melscale_fbanks128 = melscale_fbanks(n_fft // 2 + 1, 0, sample_rate // 2, 128, sample_rate, "slaney", "slaney").astype(
    np.float32
)
stft_conv_weights_nemo = stft_conv_weights(
    np.pad(np.hanning(win_length), (n_fft // 2 - win_length // 2, n_fft // 2 - win_length // 2)).astype(np.float32)
)


@script()
def normalize(x: FLOAT["batch_size", "M", "T"], lens: INT64["batch_size"]):
    lens_3d = op.Unsqueeze(lens, axes=[1, 2])
    mask = op.Range(0, op.Squeeze(op.Shape(x, start=2, end=3)), 1) < lens_3d
    lens_3d = op.CastLike(lens_3d, x)
    mean = op.ReduceSum(op.Where(mask, x, 0.0), axes=[-1], keepdims=1) / lens_3d
    var = op.ReduceSumSquare(op.Where(mask, x - mean, 0.0), axes=[-1], keepdims=1) / (lens_3d - 1)
    return op.Where(mask, (x - mean) / (op.Sqrt(var) + 1e-5), 0.0)


@script()
def nemo_preprocessor(
    waveforms: FLOAT["batch_size", "N"],
    waveforms_lens: INT64["batch_size"],
    melscale_fbanks: FLOAT[n_fft // 2 + 1, "M"],
):
    if preemph != 0.0:
        timemask = op.Range(0, op.Squeeze(op.Shape(waveforms, start=1, end=2)), 1) < op.Unsqueeze(
            waveforms_lens, axes=[1]
        )
        waveforms = op.Concat(waveforms[:, :1], waveforms[:, 1:] - preemph * waveforms[:, :-1], axis=-1)
        waveforms = op.Where(timemask, waveforms, 0.0)

    waveforms = op.Pad(
        waveforms,
        pads=op.Constant(value=[0, n_fft // 2, 0, n_fft // 2]),
    )
    hann_window = op.Pad(
        op.HannWindow(win_length, periodic=0),
        pads=op.Constant(value=[n_fft // 2 - win_length // 2, n_fft // 2 - win_length // 2]),
    )
    image = op.STFT(waveforms, hop_length, hann_window)
    spectrogram = op.ReduceSumSquare(image, axes=[-1], keepdims=0)

    mel_spectrogram = op.MatMul(spectrogram, melscale_fbanks)
    log_mel_spectrogram = op.Log(mel_spectrogram + log_zero_guard_value)

    features_lens = waveforms_lens / hop_length
    return normalize(op.Transpose(log_mel_spectrogram, perm=[0, 2, 1]), features_lens), features_lens


@script(doc_string="LogMelSpectrogram feature extractor for Nemo models", default_opset=op)
def NemoPreprocessor80(
    waveforms: FLOAT["batch_size", "N"],
    waveforms_lens: INT64["batch_size"],
) -> tuple[FLOAT["batch_size", 80, "T"], INT64["batch_size"]]:
    features, features_lens = nemo_preprocessor(
        waveforms,
        waveforms_lens,
        melscale_fbanks80,
    )
    return features, features_lens


@script(doc_string="LogMelSpectrogram feature extractor for Nemo models", default_opset=op)
def NemoPreprocessor128(
    waveforms: FLOAT["batch_size", "N"],
    waveforms_lens: INT64["batch_size"],
) -> tuple[FLOAT["batch_size", 128, "T"], INT64["batch_size"]]:
    features, features_lens = nemo_preprocessor(
        waveforms,
        waveforms_lens,
        melscale_fbanks128,
    )
    return features, features_lens


@script()
def nemo_preprocessor_conv(
    waveforms: FLOAT["batch_size", "N"],
    waveforms_lens: INT64["batch_size"],
    melscale_fbanks: FLOAT[n_fft // 2 + 1, "M"],
    conv_weights: FLOAT["channels", 1, n_fft],
):
    if preemph != 0.0:
        timemask = op.Range(0, op.Squeeze(op.Shape(waveforms, start=1, end=2)), 1) < op.Unsqueeze(
            waveforms_lens, axes=[1]
        )
        waveforms = op.Concat(waveforms[:, :1], waveforms[:, 1:] - preemph * waveforms[:, :-1], axis=-1)
        waveforms = op.Where(timemask, waveforms, 0.0)

    waveforms = op.Pad(
        waveforms,
        pads=op.Constant(value=[0, n_fft // 2, 0, n_fft // 2]),
    )
    spectrogram = conv_power_spectrogram(waveforms, conv_weights)

    mel_spectrogram = op.MatMul(spectrogram, melscale_fbanks)
    log_mel_spectrogram = op.Log(mel_spectrogram + log_zero_guard_value)

    features_lens = waveforms_lens / hop_length
    return normalize(op.Transpose(log_mel_spectrogram, perm=[0, 2, 1]), features_lens), features_lens


@script(doc_string="LogMelSpectrogram feature extractor for Nemo models (Conv-based STFT)", default_opset=op)
def NemoPreprocessor80Conv(
    waveforms: FLOAT["batch_size", "N"],
    waveforms_lens: INT64["batch_size"],
) -> tuple[FLOAT["batch_size", 80, "T"], INT64["batch_size"]]:
    features, features_lens = nemo_preprocessor_conv(
        waveforms,
        waveforms_lens,
        melscale_fbanks80,
        stft_conv_weights_nemo,
    )
    return features, features_lens


@script(doc_string="LogMelSpectrogram feature extractor for Nemo models (Conv-based STFT)", default_opset=op)
def NemoPreprocessor128Conv(
    waveforms: FLOAT["batch_size", "N"],
    waveforms_lens: INT64["batch_size"],
) -> tuple[FLOAT["batch_size", 128, "T"], INT64["batch_size"]]:
    features, features_lens = nemo_preprocessor_conv(
        waveforms,
        waveforms_lens,
        melscale_fbanks128,
        stft_conv_weights_nemo,
    )
    return features, features_lens