File size: 5,288 Bytes
71cd91e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# coding=utf-8
# Copyright 2023 The HuggingFace Inc. team and the librosa & torchaudio authors.
#
# 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.
"""
Audio processing functions to extract features from audio waveforms. This code is pure numpy to support all frameworks
and remove unnecessary dependencies.
"""
import warnings
import numpy as np
from typing import Union, Optional


def hertz_to_mel(
    freq: Union[float, np.ndarray], mel_scale: str = "htk"
) -> Union[float, np.ndarray]:
    if mel_scale not in ["slaney", "htk", "kaldi"]:
        raise ValueError('mel_scale should be one of "htk", "slaney" or "kaldi".')

    if mel_scale == "htk":
        return 2595.0 * np.log10(1.0 + (freq / 700.0))
    elif mel_scale == "kaldi":
        return 1127.0 * np.log(1.0 + (freq / 700.0))

    min_log_hertz = 1000.0
    min_log_mel = 15.0
    logstep = 27.0 / np.log(6.4)
    mels = 3.0 * freq / 200.0

    if isinstance(freq, np.ndarray):
        log_region = freq >= min_log_hertz
        mels[log_region] = (
            min_log_mel + np.log(freq[log_region] / min_log_hertz) * logstep
        )
    elif freq >= min_log_hertz:
        mels = min_log_mel + np.log(freq / min_log_hertz) * logstep

    return mels


def mel_to_hertz(
    mels: Union[float, np.ndarray], mel_scale: str = "htk"
) -> Union[float, np.ndarray]:
    if mel_scale not in ["slaney", "htk", "kaldi"]:
        raise ValueError('mel_scale should be one of "htk", "slaney" or "kaldi".')

    if mel_scale == "htk":
        return 700.0 * (np.power(10, mels / 2595.0) - 1.0)
    elif mel_scale == "kaldi":
        return 700.0 * (np.exp(mels / 1127.0) - 1.0)

    min_log_hertz = 1000.0
    min_log_mel = 15.0
    logstep = np.log(6.4) / 27.0
    freq = 200.0 * mels / 3.0

    if isinstance(mels, np.ndarray):
        log_region = mels >= min_log_mel
        freq[log_region] = min_log_hertz * np.exp(
            logstep * (mels[log_region] - min_log_mel)
        )
    elif mels >= min_log_mel:
        freq = min_log_hertz * np.exp(logstep * (mels - min_log_mel))

    return freq


def _create_triangular_filter_bank(
    fft_freqs: np.ndarray, filter_freqs: np.ndarray
) -> np.ndarray:
    """
    Creates a triangular filter bank.

    Adapted from *torchaudio* and *librosa*.

    Args:
        fft_freqs (`np.ndarray` of shape `(num_frequency_bins,)`):
            Discrete frequencies of the FFT bins in Hz.
        filter_freqs (`np.ndarray` of shape `(num_mel_filters,)`):
            Center frequencies of the triangular filters to create, in Hz.

    Returns:
        `np.ndarray` of shape `(num_frequency_bins, num_mel_filters)`
    """
    filter_diff = np.diff(filter_freqs)
    slopes = np.expand_dims(filter_freqs, 0) - np.expand_dims(fft_freqs, 1)
    down_slopes = -slopes[:, :-2] / filter_diff[:-1]
    up_slopes = slopes[:, 2:] / filter_diff[1:]
    return np.maximum(np.zeros(1), np.minimum(down_slopes, up_slopes))


def mel_filter_bank(
    num_frequency_bins: int,
    num_mel_filters: int,
    min_frequency: float,
    max_frequency: float,
    sampling_rate: int,
    norm: Optional[str] = None,
    mel_scale: str = "htk",
    triangularize_in_mel_space: bool = False,
) -> np.ndarray:
    if norm is not None and norm != "slaney":
        raise ValueError('norm must be one of None or "slaney"')

    # center points of the triangular mel filters
    mel_min = hertz_to_mel(min_frequency, mel_scale=mel_scale)
    mel_max = hertz_to_mel(max_frequency, mel_scale=mel_scale)
    mel_freqs = np.linspace(mel_min, mel_max, num_mel_filters + 2)
    filter_freqs = mel_to_hertz(mel_freqs, mel_scale=mel_scale)

    if triangularize_in_mel_space:
        # frequencies of FFT bins in Hz, but filters triangularized in mel space
        fft_bin_width = sampling_rate / (num_frequency_bins * 2)
        fft_freqs = hertz_to_mel(
            fft_bin_width * np.arange(num_frequency_bins), mel_scale=mel_scale
        )
        filter_freqs = mel_freqs
    else:
        # frequencies of FFT bins in Hz
        fft_freqs = np.linspace(0, sampling_rate // 2, num_frequency_bins)

    mel_filters = _create_triangular_filter_bank(fft_freqs, filter_freqs)

    if norm is not None and norm == "slaney":
        # Slaney-style mel is scaled to be approx constant energy per channel
        enorm = 2.0 / (
            filter_freqs[2 : num_mel_filters + 2] - filter_freqs[:num_mel_filters]
        )
        mel_filters *= np.expand_dims(enorm, 0)

    if (mel_filters.max(axis=0) == 0.0).any():
        warnings.warn(
            "At least one mel filter has all zero values. "
            f"The value for `num_mel_filters` ({num_mel_filters}) may be set too high. "
            f"Or, the value for `num_frequency_bins` ({num_frequency_bins}) may be set too low."
        )

    return mel_filters