Alex Loiko | 4ed47d0 | 2018-04-04 15:05:57 +0200 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2016 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/agc2/signal_classifier.h" |
| 12 | |
| 13 | #include <algorithm> |
| 14 | #include <numeric> |
| 15 | #include <vector> |
| 16 | |
| 17 | #include "api/array_view.h" |
| 18 | #include "modules/audio_processing/agc2/down_sampler.h" |
| 19 | #include "modules/audio_processing/agc2/noise_spectrum_estimator.h" |
| 20 | #include "modules/audio_processing/logging/apm_data_dumper.h" |
Yves Gerey | 988cc08 | 2018-10-23 12:03:01 +0200 | [diff] [blame] | 21 | #include "rtc_base/checks.h" |
Per Åhgren | 80b3012 | 2020-06-02 14:28:25 +0200 | [diff] [blame] | 22 | #include "system_wrappers/include/cpu_features_wrapper.h" |
Alex Loiko | 4ed47d0 | 2018-04-04 15:05:57 +0200 | [diff] [blame] | 23 | |
| 24 | namespace webrtc { |
| 25 | namespace { |
| 26 | |
Per Åhgren | 80b3012 | 2020-06-02 14:28:25 +0200 | [diff] [blame] | 27 | bool IsSse2Available() { |
| 28 | #if defined(WEBRTC_ARCH_X86_FAMILY) |
Mirko Bonadei | bef022b | 2020-09-06 16:07:15 +0200 | [diff] [blame] | 29 | return GetCPUInfo(kSSE2) != 0; |
Per Åhgren | 80b3012 | 2020-06-02 14:28:25 +0200 | [diff] [blame] | 30 | #else |
| 31 | return false; |
| 32 | #endif |
| 33 | } |
| 34 | |
Alex Loiko | 4ed47d0 | 2018-04-04 15:05:57 +0200 | [diff] [blame] | 35 | void RemoveDcLevel(rtc::ArrayView<float> x) { |
| 36 | RTC_DCHECK_LT(0, x.size()); |
| 37 | float mean = std::accumulate(x.data(), x.data() + x.size(), 0.f); |
| 38 | mean /= x.size(); |
| 39 | |
| 40 | for (float& v : x) { |
| 41 | v -= mean; |
| 42 | } |
| 43 | } |
| 44 | |
| 45 | void PowerSpectrum(const OouraFft* ooura_fft, |
| 46 | rtc::ArrayView<const float> x, |
| 47 | rtc::ArrayView<float> spectrum) { |
| 48 | RTC_DCHECK_EQ(65, spectrum.size()); |
| 49 | RTC_DCHECK_EQ(128, x.size()); |
| 50 | float X[128]; |
| 51 | std::copy(x.data(), x.data() + x.size(), X); |
| 52 | ooura_fft->Fft(X); |
| 53 | |
| 54 | float* X_p = X; |
| 55 | RTC_DCHECK_EQ(X_p, &X[0]); |
| 56 | spectrum[0] = (*X_p) * (*X_p); |
| 57 | ++X_p; |
| 58 | RTC_DCHECK_EQ(X_p, &X[1]); |
| 59 | spectrum[64] = (*X_p) * (*X_p); |
| 60 | for (int k = 1; k < 64; ++k) { |
| 61 | ++X_p; |
| 62 | RTC_DCHECK_EQ(X_p, &X[2 * k]); |
| 63 | spectrum[k] = (*X_p) * (*X_p); |
| 64 | ++X_p; |
| 65 | RTC_DCHECK_EQ(X_p, &X[2 * k + 1]); |
| 66 | spectrum[k] += (*X_p) * (*X_p); |
| 67 | } |
| 68 | } |
| 69 | |
| 70 | webrtc::SignalClassifier::SignalType ClassifySignal( |
| 71 | rtc::ArrayView<const float> signal_spectrum, |
| 72 | rtc::ArrayView<const float> noise_spectrum, |
| 73 | ApmDataDumper* data_dumper) { |
| 74 | int num_stationary_bands = 0; |
| 75 | int num_highly_nonstationary_bands = 0; |
| 76 | |
| 77 | // Detect stationary and highly nonstationary bands. |
| 78 | for (size_t k = 1; k < 40; k++) { |
| 79 | if (signal_spectrum[k] < 3 * noise_spectrum[k] && |
| 80 | signal_spectrum[k] * 3 > noise_spectrum[k]) { |
| 81 | ++num_stationary_bands; |
| 82 | } else if (signal_spectrum[k] > 9 * noise_spectrum[k]) { |
| 83 | ++num_highly_nonstationary_bands; |
| 84 | } |
| 85 | } |
| 86 | |
Alessio Bazzica | 70b775d | 2021-04-07 12:03:11 +0200 | [diff] [blame] | 87 | data_dumper->DumpRaw("agc2_num_stationary_bands", 1, &num_stationary_bands); |
| 88 | data_dumper->DumpRaw("agc2_num_highly_nonstationary_bands", 1, |
Alex Loiko | 4ed47d0 | 2018-04-04 15:05:57 +0200 | [diff] [blame] | 89 | &num_highly_nonstationary_bands); |
| 90 | |
| 91 | // Use the detected number of bands to classify the overall signal |
| 92 | // stationarity. |
| 93 | if (num_stationary_bands > 15) { |
| 94 | return SignalClassifier::SignalType::kStationary; |
| 95 | } else { |
| 96 | return SignalClassifier::SignalType::kNonStationary; |
| 97 | } |
| 98 | } |
| 99 | |
| 100 | } // namespace |
| 101 | |
| 102 | SignalClassifier::FrameExtender::FrameExtender(size_t frame_size, |
| 103 | size_t extended_frame_size) |
| 104 | : x_old_(extended_frame_size - frame_size, 0.f) {} |
| 105 | |
| 106 | SignalClassifier::FrameExtender::~FrameExtender() = default; |
| 107 | |
| 108 | void SignalClassifier::FrameExtender::ExtendFrame( |
| 109 | rtc::ArrayView<const float> x, |
| 110 | rtc::ArrayView<float> x_extended) { |
| 111 | RTC_DCHECK_EQ(x_old_.size() + x.size(), x_extended.size()); |
| 112 | std::copy(x_old_.data(), x_old_.data() + x_old_.size(), x_extended.data()); |
| 113 | std::copy(x.data(), x.data() + x.size(), x_extended.data() + x_old_.size()); |
| 114 | std::copy(x_extended.data() + x_extended.size() - x_old_.size(), |
| 115 | x_extended.data() + x_extended.size(), x_old_.data()); |
| 116 | } |
| 117 | |
| 118 | SignalClassifier::SignalClassifier(ApmDataDumper* data_dumper) |
| 119 | : data_dumper_(data_dumper), |
| 120 | down_sampler_(data_dumper_), |
Per Åhgren | 80b3012 | 2020-06-02 14:28:25 +0200 | [diff] [blame] | 121 | noise_spectrum_estimator_(data_dumper_), |
| 122 | ooura_fft_(IsSse2Available()) { |
Alex Loiko | 4ed47d0 | 2018-04-04 15:05:57 +0200 | [diff] [blame] | 123 | Initialize(48000); |
| 124 | } |
| 125 | SignalClassifier::~SignalClassifier() {} |
| 126 | |
| 127 | void SignalClassifier::Initialize(int sample_rate_hz) { |
| 128 | down_sampler_.Initialize(sample_rate_hz); |
| 129 | noise_spectrum_estimator_.Initialize(); |
| 130 | frame_extender_.reset(new FrameExtender(80, 128)); |
| 131 | sample_rate_hz_ = sample_rate_hz; |
| 132 | initialization_frames_left_ = 2; |
| 133 | consistent_classification_counter_ = 3; |
| 134 | last_signal_type_ = SignalClassifier::SignalType::kNonStationary; |
| 135 | } |
| 136 | |
| 137 | SignalClassifier::SignalType SignalClassifier::Analyze( |
| 138 | rtc::ArrayView<const float> signal) { |
| 139 | RTC_DCHECK_EQ(signal.size(), sample_rate_hz_ / 100); |
| 140 | |
| 141 | // Compute the signal power spectrum. |
| 142 | float downsampled_frame[80]; |
| 143 | down_sampler_.DownSample(signal, downsampled_frame); |
| 144 | float extended_frame[128]; |
| 145 | frame_extender_->ExtendFrame(downsampled_frame, extended_frame); |
| 146 | RemoveDcLevel(extended_frame); |
| 147 | float signal_spectrum[65]; |
| 148 | PowerSpectrum(&ooura_fft_, extended_frame, signal_spectrum); |
| 149 | |
| 150 | // Classify the signal based on the estimate of the noise spectrum and the |
| 151 | // signal spectrum estimate. |
| 152 | const SignalType signal_type = ClassifySignal( |
| 153 | signal_spectrum, noise_spectrum_estimator_.GetNoiseSpectrum(), |
| 154 | data_dumper_); |
| 155 | |
| 156 | // Update the noise spectrum based on the signal spectrum. |
| 157 | noise_spectrum_estimator_.Update(signal_spectrum, |
| 158 | initialization_frames_left_ > 0); |
| 159 | |
| 160 | // Update the number of frames until a reliable signal spectrum is achieved. |
| 161 | initialization_frames_left_ = std::max(0, initialization_frames_left_ - 1); |
| 162 | |
| 163 | if (last_signal_type_ == signal_type) { |
| 164 | consistent_classification_counter_ = |
| 165 | std::max(0, consistent_classification_counter_ - 1); |
| 166 | } else { |
| 167 | last_signal_type_ = signal_type; |
| 168 | consistent_classification_counter_ = 3; |
| 169 | } |
| 170 | |
| 171 | if (consistent_classification_counter_ > 0) { |
| 172 | return SignalClassifier::SignalType::kNonStationary; |
| 173 | } |
| 174 | return signal_type; |
| 175 | } |
| 176 | |
| 177 | } // namespace webrtc |