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
|
/*
* Copyright (c) 2019 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/ns/signal_model_estimator.h"
#include "modules/audio_processing/ns/fast_math.h"
namespace webrtc {
namespace {
constexpr float kOneByFftSizeBy2Plus1 = 1.f / kFftSizeBy2Plus1;
// Computes the difference measure between input spectrum and a template/learned
// noise spectrum.
float ComputeSpectralDiff(
ArrayView<const float, kFftSizeBy2Plus1> conservative_noise_spectrum,
ArrayView<const float, kFftSizeBy2Plus1> signal_spectrum,
float signal_spectral_sum,
float diff_normalization) {
// spectral_diff = var(signal_spectrum) - cov(signal_spectrum, magnAvgPause)^2
// / var(magnAvgPause)
// Compute average quantities.
float noise_average = 0.f;
for (size_t i = 0; i < kFftSizeBy2Plus1; ++i) {
// Conservative smooth noise spectrum from pause frames.
noise_average += conservative_noise_spectrum[i];
}
noise_average = noise_average * kOneByFftSizeBy2Plus1;
float signal_average = signal_spectral_sum * kOneByFftSizeBy2Plus1;
// Compute variance and covariance quantities.
float covariance = 0.f;
float noise_variance = 0.f;
float signal_variance = 0.f;
for (size_t i = 0; i < kFftSizeBy2Plus1; ++i) {
float signal_diff = signal_spectrum[i] - signal_average;
float noise_diff = conservative_noise_spectrum[i] - noise_average;
covariance += signal_diff * noise_diff;
noise_variance += noise_diff * noise_diff;
signal_variance += signal_diff * signal_diff;
}
covariance *= kOneByFftSizeBy2Plus1;
noise_variance *= kOneByFftSizeBy2Plus1;
signal_variance *= kOneByFftSizeBy2Plus1;
// Update of average magnitude spectrum.
float spectral_diff =
signal_variance - (covariance * covariance) / (noise_variance + 0.0001f);
// Normalize.
return spectral_diff / (diff_normalization + 0.0001f);
}
// Updates the spectral flatness based on the input spectrum.
void UpdateSpectralFlatness(
ArrayView<const float, kFftSizeBy2Plus1> signal_spectrum,
float signal_spectral_sum,
float* spectral_flatness) {
RTC_DCHECK(spectral_flatness);
// Compute log of ratio of the geometric to arithmetic mean (handle the log(0)
// separately).
constexpr float kAveraging = 0.3f;
float avg_spect_flatness_num = 0.f;
for (size_t i = 1; i < kFftSizeBy2Plus1; ++i) {
if (signal_spectrum[i] == 0.f) {
*spectral_flatness -= kAveraging * (*spectral_flatness);
return;
}
}
for (size_t i = 1; i < kFftSizeBy2Plus1; ++i) {
avg_spect_flatness_num += LogApproximation(signal_spectrum[i]);
}
float avg_spect_flatness_denom = signal_spectral_sum - signal_spectrum[0];
avg_spect_flatness_denom = avg_spect_flatness_denom * kOneByFftSizeBy2Plus1;
avg_spect_flatness_num = avg_spect_flatness_num * kOneByFftSizeBy2Plus1;
float spectral_tmp =
ExpApproximation(avg_spect_flatness_num) / avg_spect_flatness_denom;
// Time-avg update of spectral flatness feature.
*spectral_flatness += kAveraging * (spectral_tmp - *spectral_flatness);
}
// Updates the log LRT measures.
void UpdateSpectralLrt(ArrayView<const float, kFftSizeBy2Plus1> prior_snr,
ArrayView<const float, kFftSizeBy2Plus1> post_snr,
ArrayView<float, kFftSizeBy2Plus1> avg_log_lrt,
float* lrt) {
RTC_DCHECK(lrt);
for (size_t i = 0; i < kFftSizeBy2Plus1; ++i) {
float tmp1 = 1.f + 2.f * prior_snr[i];
float tmp2 = 2.f * prior_snr[i] / (tmp1 + 0.0001f);
float bessel_tmp = (post_snr[i] + 1.f) * tmp2;
avg_log_lrt[i] +=
.5f * (bessel_tmp - LogApproximation(tmp1) - avg_log_lrt[i]);
}
float log_lrt_time_avg_k_sum = 0.f;
for (size_t i = 0; i < kFftSizeBy2Plus1; ++i) {
log_lrt_time_avg_k_sum += avg_log_lrt[i];
}
*lrt = log_lrt_time_avg_k_sum * kOneByFftSizeBy2Plus1;
}
} // namespace
SignalModelEstimator::SignalModelEstimator()
: prior_model_estimator_(kLtrFeatureThr) {}
void SignalModelEstimator::AdjustNormalization(int32_t num_analyzed_frames,
float signal_energy) {
diff_normalization_ *= num_analyzed_frames;
diff_normalization_ += signal_energy;
diff_normalization_ /= (num_analyzed_frames + 1);
}
// Update the noise features.
void SignalModelEstimator::Update(
ArrayView<const float, kFftSizeBy2Plus1> prior_snr,
ArrayView<const float, kFftSizeBy2Plus1> post_snr,
ArrayView<const float, kFftSizeBy2Plus1> conservative_noise_spectrum,
ArrayView<const float, kFftSizeBy2Plus1> signal_spectrum,
float signal_spectral_sum,
float signal_energy) {
// Compute spectral flatness on input spectrum.
UpdateSpectralFlatness(signal_spectrum, signal_spectral_sum,
&features_.spectral_flatness);
// Compute difference of input spectrum with learned/estimated noise spectrum.
float spectral_diff =
ComputeSpectralDiff(conservative_noise_spectrum, signal_spectrum,
signal_spectral_sum, diff_normalization_);
// Compute time-avg update of difference feature.
features_.spectral_diff += 0.3f * (spectral_diff - features_.spectral_diff);
signal_energy_sum_ += signal_energy;
// Compute histograms for parameter decisions (thresholds and weights for
// features). Parameters are extracted periodically.
if (--histogram_analysis_counter_ > 0) {
histograms_.Update(features_);
} else {
// Compute model parameters.
prior_model_estimator_.Update(histograms_);
// Clear histograms for next update.
histograms_.Clear();
histogram_analysis_counter_ = kFeatureUpdateWindowSize;
// Update every window:
// Compute normalization for the spectral difference for next estimation.
signal_energy_sum_ = signal_energy_sum_ / kFeatureUpdateWindowSize;
diff_normalization_ = 0.5f * (signal_energy_sum_ + diff_normalization_);
signal_energy_sum_ = 0.f;
}
// Compute the LRT.
UpdateSpectralLrt(prior_snr, post_snr, features_.avg_log_lrt, &features_.lrt);
}
} // namespace webrtc
|
