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
|
/*
* Copyright (c) 2018 The WebRTC project authors. All Rights Reserved.
*
* Use of this source code is governed by a BSD-style license
* that can be found in the LICENSE file in the root of the source
* tree. An additional intellectual property rights grant can be found
* in the file PATENTS. All contributing project authors may
* be found in the AUTHORS file in the root of the source tree.
*/
#include "modules/audio_processing/agc2/rnn_vad/spectral_features.h"
#include <algorithm>
#include <cmath>
#include <limits>
#include <numeric>
#include "rtc_base/checks.h"
#include "rtc_base/numerics/safe_compare.h"
namespace webrtc {
namespace rnn_vad {
namespace {
constexpr float kSilenceThreshold = 0.04f;
// Computes the new cepstral difference stats and pushes them into the passed
// symmetric matrix buffer.
void UpdateCepstralDifferenceStats(
ArrayView<const float, kNumBands> new_cepstral_coeffs,
const RingBuffer<float, kNumBands, kCepstralCoeffsHistorySize>& ring_buf,
SymmetricMatrixBuffer<float, kCepstralCoeffsHistorySize>* sym_matrix_buf) {
RTC_DCHECK(sym_matrix_buf);
// Compute the new cepstral distance stats.
std::array<float, kCepstralCoeffsHistorySize - 1> distances;
for (int i = 0; i < kCepstralCoeffsHistorySize - 1; ++i) {
const int delay = i + 1;
auto old_cepstral_coeffs = ring_buf.GetArrayView(delay);
distances[i] = 0.f;
for (int k = 0; k < kNumBands; ++k) {
const float c = new_cepstral_coeffs[k] - old_cepstral_coeffs[k];
distances[i] += c * c;
}
}
// Push the new spectral distance stats into the symmetric matrix buffer.
sym_matrix_buf->Push(distances);
}
// Computes the first half of the Vorbis window.
std::array<float, kFrameSize20ms24kHz / 2> ComputeScaledHalfVorbisWindow(
float scaling = 1.f) {
constexpr int kHalfSize = kFrameSize20ms24kHz / 2;
std::array<float, kHalfSize> half_window{};
for (int i = 0; i < kHalfSize; ++i) {
half_window[i] =
scaling *
std::sin(0.5 * kPi * std::sin(0.5 * kPi * (i + 0.5) / kHalfSize) *
std::sin(0.5 * kPi * (i + 0.5) / kHalfSize));
}
return half_window;
}
// Computes the forward FFT on a 20 ms frame to which a given window function is
// applied. The Fourier coefficient corresponding to the Nyquist frequency is
// set to zero (it is never used and this allows to simplify the code).
void ComputeWindowedForwardFft(
ArrayView<const float, kFrameSize20ms24kHz> frame,
const std::array<float, kFrameSize20ms24kHz / 2>& half_window,
Pffft::FloatBuffer* fft_input_buffer,
Pffft::FloatBuffer* fft_output_buffer,
Pffft* fft) {
RTC_DCHECK_EQ(frame.size(), 2 * half_window.size());
// Apply windowing.
auto in = fft_input_buffer->GetView();
for (int i = 0, j = kFrameSize20ms24kHz - 1; SafeLt(i, half_window.size());
++i, --j) {
in[i] = frame[i] * half_window[i];
in[j] = frame[j] * half_window[i];
}
fft->ForwardTransform(*fft_input_buffer, fft_output_buffer, /*ordered=*/true);
// Set the Nyquist frequency coefficient to zero.
auto out = fft_output_buffer->GetView();
out[1] = 0.f;
}
} // namespace
SpectralFeaturesExtractor::SpectralFeaturesExtractor()
: half_window_(ComputeScaledHalfVorbisWindow(
1.f / static_cast<float>(kFrameSize20ms24kHz))),
fft_(kFrameSize20ms24kHz, Pffft::FftType::kReal),
fft_buffer_(fft_.CreateBuffer()),
reference_frame_fft_(fft_.CreateBuffer()),
lagged_frame_fft_(fft_.CreateBuffer()),
dct_table_(ComputeDctTable()) {}
SpectralFeaturesExtractor::~SpectralFeaturesExtractor() = default;
void SpectralFeaturesExtractor::Reset() {
cepstral_coeffs_ring_buf_.Reset();
cepstral_diffs_buf_.Reset();
}
bool SpectralFeaturesExtractor::CheckSilenceComputeFeatures(
ArrayView<const float, kFrameSize20ms24kHz> reference_frame,
ArrayView<const float, kFrameSize20ms24kHz> lagged_frame,
ArrayView<float, kNumBands - kNumLowerBands> higher_bands_cepstrum,
ArrayView<float, kNumLowerBands> average,
ArrayView<float, kNumLowerBands> first_derivative,
ArrayView<float, kNumLowerBands> second_derivative,
ArrayView<float, kNumLowerBands> bands_cross_corr,
float* variability) {
// Compute the Opus band energies for the reference frame.
ComputeWindowedForwardFft(reference_frame, half_window_, fft_buffer_.get(),
reference_frame_fft_.get(), &fft_);
spectral_correlator_.ComputeAutoCorrelation(
reference_frame_fft_->GetConstView(), reference_frame_bands_energy_);
// Check if the reference frame has silence.
const float tot_energy =
std::accumulate(reference_frame_bands_energy_.begin(),
reference_frame_bands_energy_.end(), 0.f);
if (tot_energy < kSilenceThreshold) {
return true;
}
// Compute the Opus band energies for the lagged frame.
ComputeWindowedForwardFft(lagged_frame, half_window_, fft_buffer_.get(),
lagged_frame_fft_.get(), &fft_);
spectral_correlator_.ComputeAutoCorrelation(lagged_frame_fft_->GetConstView(),
lagged_frame_bands_energy_);
// Log of the band energies for the reference frame.
std::array<float, kNumBands> log_bands_energy;
ComputeSmoothedLogMagnitudeSpectrum(reference_frame_bands_energy_,
log_bands_energy);
// Reference frame cepstrum.
std::array<float, kNumBands> cepstrum;
ComputeDct(log_bands_energy, dct_table_, cepstrum);
// Ad-hoc correction terms for the first two cepstral coefficients.
cepstrum[0] -= 12.f;
cepstrum[1] -= 4.f;
// Update the ring buffer and the cepstral difference stats.
cepstral_coeffs_ring_buf_.Push(cepstrum);
UpdateCepstralDifferenceStats(cepstrum, cepstral_coeffs_ring_buf_,
&cepstral_diffs_buf_);
// Write the higher bands cepstral coefficients.
RTC_DCHECK_EQ(cepstrum.size() - kNumLowerBands, higher_bands_cepstrum.size());
std::copy(cepstrum.begin() + kNumLowerBands, cepstrum.end(),
higher_bands_cepstrum.begin());
// Compute and write remaining features.
ComputeAvgAndDerivatives(average, first_derivative, second_derivative);
ComputeNormalizedCepstralCorrelation(bands_cross_corr);
RTC_DCHECK(variability);
*variability = ComputeVariability();
return false;
}
void SpectralFeaturesExtractor::ComputeAvgAndDerivatives(
ArrayView<float, kNumLowerBands> average,
ArrayView<float, kNumLowerBands> first_derivative,
ArrayView<float, kNumLowerBands> second_derivative) const {
auto curr = cepstral_coeffs_ring_buf_.GetArrayView(0);
auto prev1 = cepstral_coeffs_ring_buf_.GetArrayView(1);
auto prev2 = cepstral_coeffs_ring_buf_.GetArrayView(2);
RTC_DCHECK_EQ(average.size(), first_derivative.size());
RTC_DCHECK_EQ(first_derivative.size(), second_derivative.size());
RTC_DCHECK_LE(average.size(), curr.size());
for (int i = 0; SafeLt(i, average.size()); ++i) {
// Average, kernel: [1, 1, 1].
average[i] = curr[i] + prev1[i] + prev2[i];
// First derivative, kernel: [1, 0, - 1].
first_derivative[i] = curr[i] - prev2[i];
// Second derivative, Laplacian kernel: [1, -2, 1].
second_derivative[i] = curr[i] - 2 * prev1[i] + prev2[i];
}
}
void SpectralFeaturesExtractor::ComputeNormalizedCepstralCorrelation(
ArrayView<float, kNumLowerBands> bands_cross_corr) {
spectral_correlator_.ComputeCrossCorrelation(
reference_frame_fft_->GetConstView(), lagged_frame_fft_->GetConstView(),
bands_cross_corr_);
// Normalize.
for (int i = 0; SafeLt(i, bands_cross_corr_.size()); ++i) {
bands_cross_corr_[i] =
bands_cross_corr_[i] /
std::sqrt(0.001f + reference_frame_bands_energy_[i] *
lagged_frame_bands_energy_[i]);
}
// Cepstrum.
ComputeDct(bands_cross_corr_, dct_table_, bands_cross_corr);
// Ad-hoc correction terms for the first two cepstral coefficients.
bands_cross_corr[0] -= 1.3f;
bands_cross_corr[1] -= 0.9f;
}
float SpectralFeaturesExtractor::ComputeVariability() const {
// Compute cepstral variability score.
float variability = 0.f;
for (int delay1 = 0; delay1 < kCepstralCoeffsHistorySize; ++delay1) {
float min_dist = std::numeric_limits<float>::max();
for (int delay2 = 0; delay2 < kCepstralCoeffsHistorySize; ++delay2) {
if (delay1 == delay2) // The distance would be 0.
continue;
min_dist =
std::min(min_dist, cepstral_diffs_buf_.GetValue(delay1, delay2));
}
variability += min_dist;
}
// Normalize (based on training set stats).
// TODO(bugs.webrtc.org/10480): Isolate normalization from feature extraction.
return variability / kCepstralCoeffsHistorySize - 2.1f;
}
} // namespace rnn_vad
} // namespace webrtc
|
