1 /*
2 * Copyright (c) 2019 The WebRTC project authors. All Rights Reserved.
3 *
4 * Use of this source code is governed by a BSD-style license
5 * that can be found in the LICENSE file in the root of the source
6 * tree. An additional intellectual property rights grant can be found
7 * in the file PATENTS. All contributing project authors may
8 * be found in the AUTHORS file in the root of the source tree.
9 */
10
11 #include "modules/audio_processing/ns/prior_signal_model_estimator.h"
12
13 #include <math.h>
14 #include <algorithm>
15
16 #include "modules/audio_processing/ns/fast_math.h"
17 #include "rtc_base/checks.h"
18
19 namespace webrtc {
20
21 namespace {
22
23 // Identifies the first of the two largest peaks in the histogram.
FindFirstOfTwoLargestPeaks(float bin_size,rtc::ArrayView<const int,kHistogramSize> spectral_flatness,float * peak_position,int * peak_weight)24 void FindFirstOfTwoLargestPeaks(
25 float bin_size,
26 rtc::ArrayView<const int, kHistogramSize> spectral_flatness,
27 float* peak_position,
28 int* peak_weight) {
29 RTC_DCHECK(peak_position);
30 RTC_DCHECK(peak_weight);
31
32 int peak_value = 0;
33 int secondary_peak_value = 0;
34 *peak_position = 0.f;
35 float secondary_peak_position = 0.f;
36 *peak_weight = 0;
37 int secondary_peak_weight = 0;
38
39 // Identify the two largest peaks.
40 for (int i = 0; i < kHistogramSize; ++i) {
41 const float bin_mid = (i + 0.5f) * bin_size;
42 if (spectral_flatness[i] > peak_value) {
43 // Found new "first" peak candidate.
44 secondary_peak_value = peak_value;
45 secondary_peak_weight = *peak_weight;
46 secondary_peak_position = *peak_position;
47
48 peak_value = spectral_flatness[i];
49 *peak_weight = spectral_flatness[i];
50 *peak_position = bin_mid;
51 } else if (spectral_flatness[i] > secondary_peak_value) {
52 // Found new "second" peak candidate.
53 secondary_peak_value = spectral_flatness[i];
54 secondary_peak_weight = spectral_flatness[i];
55 secondary_peak_position = bin_mid;
56 }
57 }
58
59 // Merge the peaks if they are close.
60 if ((fabs(secondary_peak_position - *peak_position) < 2 * bin_size) &&
61 (secondary_peak_weight > 0.5f * (*peak_weight))) {
62 *peak_weight += secondary_peak_weight;
63 *peak_position = 0.5f * (*peak_position + secondary_peak_position);
64 }
65 }
66
UpdateLrt(rtc::ArrayView<const int,kHistogramSize> lrt_histogram,float * prior_model_lrt,bool * low_lrt_fluctuations)67 void UpdateLrt(rtc::ArrayView<const int, kHistogramSize> lrt_histogram,
68 float* prior_model_lrt,
69 bool* low_lrt_fluctuations) {
70 RTC_DCHECK(prior_model_lrt);
71 RTC_DCHECK(low_lrt_fluctuations);
72
73 float average = 0.f;
74 float average_compl = 0.f;
75 float average_squared = 0.f;
76 int count = 0;
77
78 for (int i = 0; i < 10; ++i) {
79 float bin_mid = (i + 0.5f) * kBinSizeLrt;
80 average += lrt_histogram[i] * bin_mid;
81 count += lrt_histogram[i];
82 }
83 if (count > 0) {
84 average = average / count;
85 }
86
87 for (int i = 0; i < kHistogramSize; ++i) {
88 float bin_mid = (i + 0.5f) * kBinSizeLrt;
89 average_squared += lrt_histogram[i] * bin_mid * bin_mid;
90 average_compl += lrt_histogram[i] * bin_mid;
91 }
92 constexpr float kOneFeatureUpdateWindowSize = 1.f / kFeatureUpdateWindowSize;
93 average_squared = average_squared * kOneFeatureUpdateWindowSize;
94 average_compl = average_compl * kOneFeatureUpdateWindowSize;
95
96 // Fluctuation limit of LRT feature.
97 *low_lrt_fluctuations = average_squared - average * average_compl < 0.05f;
98
99 // Get threshold for LRT feature.
100 constexpr float kMaxLrt = 1.f;
101 constexpr float kMinLrt = .2f;
102 if (*low_lrt_fluctuations) {
103 // Very low fluctuation, so likely noise.
104 *prior_model_lrt = kMaxLrt;
105 } else {
106 *prior_model_lrt = std::min(kMaxLrt, std::max(kMinLrt, 1.2f * average));
107 }
108 }
109
110 } // namespace
111
PriorSignalModelEstimator(float lrt_initial_value)112 PriorSignalModelEstimator::PriorSignalModelEstimator(float lrt_initial_value)
113 : prior_model_(lrt_initial_value) {}
114
115 // Extract thresholds for feature parameters and computes the threshold/weights.
Update(const Histograms & histograms)116 void PriorSignalModelEstimator::Update(const Histograms& histograms) {
117 bool low_lrt_fluctuations;
118 UpdateLrt(histograms.get_lrt(), &prior_model_.lrt, &low_lrt_fluctuations);
119
120 // For spectral flatness and spectral difference: compute the main peaks of
121 // the histograms.
122 float spectral_flatness_peak_position;
123 int spectral_flatness_peak_weight;
124 FindFirstOfTwoLargestPeaks(
125 kBinSizeSpecFlat, histograms.get_spectral_flatness(),
126 &spectral_flatness_peak_position, &spectral_flatness_peak_weight);
127
128 float spectral_diff_peak_position = 0.f;
129 int spectral_diff_peak_weight = 0;
130 FindFirstOfTwoLargestPeaks(kBinSizeSpecDiff, histograms.get_spectral_diff(),
131 &spectral_diff_peak_position,
132 &spectral_diff_peak_weight);
133
134 // Reject if weight of peaks is not large enough, or peak value too small.
135 // Peak limit for spectral flatness (varies between 0 and 1).
136 const int use_spec_flat = spectral_flatness_peak_weight < 0.3f * 500 ||
137 spectral_flatness_peak_position < 0.6f
138 ? 0
139 : 1;
140
141 // Reject if weight of peaks is not large enough or if fluctuation of the LRT
142 // feature are very low, indicating a noise state.
143 const int use_spec_diff =
144 spectral_diff_peak_weight < 0.3f * 500 || low_lrt_fluctuations ? 0 : 1;
145
146 // Update the model.
147 prior_model_.template_diff_threshold = 1.2f * spectral_diff_peak_position;
148 prior_model_.template_diff_threshold =
149 std::min(1.f, std::max(0.16f, prior_model_.template_diff_threshold));
150
151 float one_by_feature_sum = 1.f / (1.f + use_spec_flat + use_spec_diff);
152 prior_model_.lrt_weighting = one_by_feature_sum;
153
154 if (use_spec_flat == 1) {
155 prior_model_.flatness_threshold = 0.9f * spectral_flatness_peak_position;
156 prior_model_.flatness_threshold =
157 std::min(.95f, std::max(0.1f, prior_model_.flatness_threshold));
158 prior_model_.flatness_weighting = one_by_feature_sum;
159 } else {
160 prior_model_.flatness_weighting = 0.f;
161 }
162
163 if (use_spec_diff == 1) {
164 prior_model_.difference_weighting = one_by_feature_sum;
165 } else {
166 prior_model_.difference_weighting = 0.f;
167 }
168 }
169
170 } // namespace webrtc
171