blob: 6fce858f162e01ef67e86ae78b2464d82e57c999 [file] [log] [blame]
/*
* Copyright (c) 2017 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/aec3/refined_filter_update_gain.h"
#include <algorithm>
#include <numeric>
#include <string>
#include <vector>
#include "modules/audio_processing/aec3/adaptive_fir_filter.h"
#include "modules/audio_processing/aec3/adaptive_fir_filter_erl.h"
#include "modules/audio_processing/aec3/aec_state.h"
#include "modules/audio_processing/aec3/coarse_filter_update_gain.h"
#include "modules/audio_processing/aec3/render_delay_buffer.h"
#include "modules/audio_processing/aec3/render_signal_analyzer.h"
#include "modules/audio_processing/aec3/subtractor_output.h"
#include "modules/audio_processing/logging/apm_data_dumper.h"
#include "modules/audio_processing/test/echo_canceller_test_tools.h"
#include "rtc_base/numerics/safe_minmax.h"
#include "rtc_base/random.h"
#include "rtc_base/strings/string_builder.h"
#include "test/gtest.h"
namespace webrtc {
namespace {
// Method for performing the simulations needed to test the refined filter
// update gain functionality.
void RunFilterUpdateTest(int num_blocks_to_process,
size_t delay_samples,
int filter_length_blocks,
const std::vector<int>& blocks_with_echo_path_changes,
const std::vector<int>& blocks_with_saturation,
bool use_silent_render_in_second_half,
std::array<float, kBlockSize>* e_last_block,
std::array<float, kBlockSize>* y_last_block,
FftData* G_last_block) {
ApmDataDumper data_dumper(42);
Aec3Optimization optimization = DetectOptimization();
constexpr size_t kNumRenderChannels = 1;
constexpr size_t kNumCaptureChannels = 1;
constexpr int kSampleRateHz = 48000;
constexpr size_t kNumBands = NumBandsForRate(kSampleRateHz);
EchoCanceller3Config config;
config.filter.refined.length_blocks = filter_length_blocks;
config.filter.coarse.length_blocks = filter_length_blocks;
AdaptiveFirFilter refined_filter(
config.filter.refined.length_blocks, config.filter.refined.length_blocks,
config.filter.config_change_duration_blocks, kNumRenderChannels,
optimization, &data_dumper);
AdaptiveFirFilter coarse_filter(
config.filter.coarse.length_blocks, config.filter.coarse.length_blocks,
config.filter.config_change_duration_blocks, kNumRenderChannels,
optimization, &data_dumper);
std::vector<std::vector<std::array<float, kFftLengthBy2Plus1>>> H2(
kNumCaptureChannels, std::vector<std::array<float, kFftLengthBy2Plus1>>(
refined_filter.max_filter_size_partitions(),
std::array<float, kFftLengthBy2Plus1>()));
for (auto& H2_ch : H2) {
for (auto& H2_k : H2_ch) {
H2_k.fill(0.f);
}
}
std::vector<std::vector<float>> h(
kNumCaptureChannels,
std::vector<float>(
GetTimeDomainLength(refined_filter.max_filter_size_partitions()),
0.f));
Aec3Fft fft;
std::array<float, kBlockSize> x_old;
x_old.fill(0.f);
CoarseFilterUpdateGain coarse_gain(
config.filter.coarse, config.filter.config_change_duration_blocks);
RefinedFilterUpdateGain refined_gain(
config.filter.refined, config.filter.config_change_duration_blocks);
Random random_generator(42U);
std::vector<std::vector<std::vector<float>>> x(
kNumBands, std::vector<std::vector<float>>(
kNumRenderChannels, std::vector<float>(kBlockSize, 0.f)));
std::vector<float> y(kBlockSize, 0.f);
config.delay.default_delay = 1;
std::unique_ptr<RenderDelayBuffer> render_delay_buffer(
RenderDelayBuffer::Create(config, kSampleRateHz, kNumRenderChannels));
AecState aec_state(config, kNumCaptureChannels);
RenderSignalAnalyzer render_signal_analyzer(config);
absl::optional<DelayEstimate> delay_estimate;
std::array<float, kFftLength> s_scratch;
std::array<float, kBlockSize> s;
FftData S;
FftData G;
std::vector<SubtractorOutput> output(kNumCaptureChannels);
for (auto& subtractor_output : output) {
subtractor_output.Reset();
}
FftData& E_refined = output[0].E_refined;
FftData E_coarse;
std::vector<std::array<float, kFftLengthBy2Plus1>> Y2(kNumCaptureChannels);
std::vector<std::array<float, kFftLengthBy2Plus1>> E2_refined(
kNumCaptureChannels);
std::array<float, kBlockSize>& e_refined = output[0].e_refined;
std::array<float, kBlockSize>& e_coarse = output[0].e_coarse;
for (auto& Y2_ch : Y2) {
Y2_ch.fill(0.f);
}
constexpr float kScale = 1.0f / kFftLengthBy2;
DelayBuffer<float> delay_buffer(delay_samples);
for (int k = 0; k < num_blocks_to_process; ++k) {
// Handle echo path changes.
if (std::find(blocks_with_echo_path_changes.begin(),
blocks_with_echo_path_changes.end(),
k) != blocks_with_echo_path_changes.end()) {
refined_filter.HandleEchoPathChange();
}
// Handle saturation.
const bool saturation =
std::find(blocks_with_saturation.begin(), blocks_with_saturation.end(),
k) != blocks_with_saturation.end();
// Create the render signal.
if (use_silent_render_in_second_half && k > num_blocks_to_process / 2) {
for (size_t band = 0; band < x.size(); ++band) {
for (size_t channel = 0; channel < x[band].size(); ++channel) {
std::fill(x[band][channel].begin(), x[band][channel].end(), 0.f);
}
}
} else {
for (size_t band = 0; band < x.size(); ++band) {
for (size_t channel = 0; channel < x[band].size(); ++channel) {
RandomizeSampleVector(&random_generator, x[band][channel]);
}
}
}
delay_buffer.Delay(x[0][0], y);
render_delay_buffer->Insert(x);
if (k == 0) {
render_delay_buffer->Reset();
}
render_delay_buffer->PrepareCaptureProcessing();
render_signal_analyzer.Update(*render_delay_buffer->GetRenderBuffer(),
aec_state.MinDirectPathFilterDelay());
// Apply the refined filter.
refined_filter.Filter(*render_delay_buffer->GetRenderBuffer(), &S);
fft.Ifft(S, &s_scratch);
std::transform(y.begin(), y.end(), s_scratch.begin() + kFftLengthBy2,
e_refined.begin(),
[&](float a, float b) { return a - b * kScale; });
std::for_each(e_refined.begin(), e_refined.end(),
[](float& a) { a = rtc::SafeClamp(a, -32768.f, 32767.f); });
fft.ZeroPaddedFft(e_refined, Aec3Fft::Window::kRectangular, &E_refined);
for (size_t k = 0; k < kBlockSize; ++k) {
s[k] = kScale * s_scratch[k + kFftLengthBy2];
}
// Apply the coarse filter.
coarse_filter.Filter(*render_delay_buffer->GetRenderBuffer(), &S);
fft.Ifft(S, &s_scratch);
std::transform(y.begin(), y.end(), s_scratch.begin() + kFftLengthBy2,
e_coarse.begin(),
[&](float a, float b) { return a - b * kScale; });
std::for_each(e_coarse.begin(), e_coarse.end(),
[](float& a) { a = rtc::SafeClamp(a, -32768.f, 32767.f); });
fft.ZeroPaddedFft(e_coarse, Aec3Fft::Window::kRectangular, &E_coarse);
// Compute spectra for future use.
E_refined.Spectrum(Aec3Optimization::kNone, output[0].E2_refined);
E_coarse.Spectrum(Aec3Optimization::kNone, output[0].E2_coarse);
// Adapt the coarse filter.
std::array<float, kFftLengthBy2Plus1> render_power;
render_delay_buffer->GetRenderBuffer()->SpectralSum(
coarse_filter.SizePartitions(), &render_power);
coarse_gain.Compute(render_power, render_signal_analyzer, E_coarse,
coarse_filter.SizePartitions(), saturation, &G);
coarse_filter.Adapt(*render_delay_buffer->GetRenderBuffer(), G);
// Adapt the refined filter
render_delay_buffer->GetRenderBuffer()->SpectralSum(
refined_filter.SizePartitions(), &render_power);
std::array<float, kFftLengthBy2Plus1> erl;
ComputeErl(optimization, H2[0], erl);
refined_gain.Compute(render_power, render_signal_analyzer, output[0], erl,
refined_filter.SizePartitions(), saturation, false,
&G);
refined_filter.Adapt(*render_delay_buffer->GetRenderBuffer(), G, &h[0]);
// Update the delay.
aec_state.HandleEchoPathChange(EchoPathVariability(
false, EchoPathVariability::DelayAdjustment::kNone, false));
refined_filter.ComputeFrequencyResponse(&H2[0]);
std::copy(output[0].E2_refined.begin(), output[0].E2_refined.end(),
E2_refined[0].begin());
aec_state.Update(delay_estimate, H2, h,
*render_delay_buffer->GetRenderBuffer(), E2_refined, Y2,
output);
}
std::copy(e_refined.begin(), e_refined.end(), e_last_block->begin());
std::copy(y.begin(), y.end(), y_last_block->begin());
std::copy(G.re.begin(), G.re.end(), G_last_block->re.begin());
std::copy(G.im.begin(), G.im.end(), G_last_block->im.begin());
}
std::string ProduceDebugText(int filter_length_blocks) {
rtc::StringBuilder ss;
ss << "Length: " << filter_length_blocks;
return ss.Release();
}
std::string ProduceDebugText(size_t delay, int filter_length_blocks) {
rtc::StringBuilder ss;
ss << "Delay: " << delay << ", ";
ss << ProduceDebugText(filter_length_blocks);
return ss.Release();
}
} // namespace
#if RTC_DCHECK_IS_ON && GTEST_HAS_DEATH_TEST && !defined(WEBRTC_ANDROID)
// Verifies that the check for non-null output gain parameter works.
TEST(RefinedFilterUpdateGainDeathTest, NullDataOutputGain) {
ApmDataDumper data_dumper(42);
EchoCanceller3Config config;
RenderSignalAnalyzer analyzer(config);
SubtractorOutput output;
RefinedFilterUpdateGain gain(config.filter.refined,
config.filter.config_change_duration_blocks);
std::array<float, kFftLengthBy2Plus1> render_power;
render_power.fill(0.f);
std::array<float, kFftLengthBy2Plus1> erl;
erl.fill(0.f);
EXPECT_DEATH(
gain.Compute(render_power, analyzer, output, erl,
config.filter.refined.length_blocks, false, false, nullptr),
"");
}
#endif
// Verifies that the gain formed causes the filter using it to converge.
TEST(RefinedFilterUpdateGain, GainCausesFilterToConverge) {
std::vector<int> blocks_with_echo_path_changes;
std::vector<int> blocks_with_saturation;
for (size_t filter_length_blocks : {12, 20, 30}) {
for (size_t delay_samples : {0, 64, 150, 200, 301}) {
SCOPED_TRACE(ProduceDebugText(delay_samples, filter_length_blocks));
std::array<float, kBlockSize> e;
std::array<float, kBlockSize> y;
FftData G;
RunFilterUpdateTest(600, delay_samples, filter_length_blocks,
blocks_with_echo_path_changes, blocks_with_saturation,
false, &e, &y, &G);
// Verify that the refined filter is able to perform well.
// Use different criteria to take overmodelling into account.
if (filter_length_blocks == 12) {
EXPECT_LT(1000 * std::inner_product(e.begin(), e.end(), e.begin(), 0.f),
std::inner_product(y.begin(), y.end(), y.begin(), 0.f));
} else {
EXPECT_LT(std::inner_product(e.begin(), e.end(), e.begin(), 0.f),
std::inner_product(y.begin(), y.end(), y.begin(), 0.f));
}
}
}
}
// Verifies that the magnitude of the gain on average decreases for a
// persistently exciting signal.
TEST(RefinedFilterUpdateGain, DecreasingGain) {
std::vector<int> blocks_with_echo_path_changes;
std::vector<int> blocks_with_saturation;
std::array<float, kBlockSize> e;
std::array<float, kBlockSize> y;
FftData G_a;
FftData G_b;
FftData G_c;
std::array<float, kFftLengthBy2Plus1> G_a_power;
std::array<float, kFftLengthBy2Plus1> G_b_power;
std::array<float, kFftLengthBy2Plus1> G_c_power;
RunFilterUpdateTest(250, 65, 12, blocks_with_echo_path_changes,
blocks_with_saturation, false, &e, &y, &G_a);
RunFilterUpdateTest(500, 65, 12, blocks_with_echo_path_changes,
blocks_with_saturation, false, &e, &y, &G_b);
RunFilterUpdateTest(750, 65, 12, blocks_with_echo_path_changes,
blocks_with_saturation, false, &e, &y, &G_c);
G_a.Spectrum(Aec3Optimization::kNone, G_a_power);
G_b.Spectrum(Aec3Optimization::kNone, G_b_power);
G_c.Spectrum(Aec3Optimization::kNone, G_c_power);
EXPECT_GT(std::accumulate(G_a_power.begin(), G_a_power.end(), 0.),
std::accumulate(G_b_power.begin(), G_b_power.end(), 0.));
EXPECT_GT(std::accumulate(G_b_power.begin(), G_b_power.end(), 0.),
std::accumulate(G_c_power.begin(), G_c_power.end(), 0.));
}
// Verifies that the gain is zero when there is saturation and that the internal
// error estimates cause the gain to increase after a period of saturation.
TEST(RefinedFilterUpdateGain, SaturationBehavior) {
std::vector<int> blocks_with_echo_path_changes;
std::vector<int> blocks_with_saturation;
for (int k = 99; k < 200; ++k) {
blocks_with_saturation.push_back(k);
}
for (size_t filter_length_blocks : {12, 20, 30}) {
SCOPED_TRACE(ProduceDebugText(filter_length_blocks));
std::array<float, kBlockSize> e;
std::array<float, kBlockSize> y;
FftData G_a;
FftData G_b;
FftData G_a_ref;
G_a_ref.re.fill(0.f);
G_a_ref.im.fill(0.f);
std::array<float, kFftLengthBy2Plus1> G_a_power;
std::array<float, kFftLengthBy2Plus1> G_b_power;
RunFilterUpdateTest(100, 65, filter_length_blocks,
blocks_with_echo_path_changes, blocks_with_saturation,
false, &e, &y, &G_a);
EXPECT_EQ(G_a_ref.re, G_a.re);
EXPECT_EQ(G_a_ref.im, G_a.im);
RunFilterUpdateTest(99, 65, filter_length_blocks,
blocks_with_echo_path_changes, blocks_with_saturation,
false, &e, &y, &G_a);
RunFilterUpdateTest(201, 65, filter_length_blocks,
blocks_with_echo_path_changes, blocks_with_saturation,
false, &e, &y, &G_b);
G_a.Spectrum(Aec3Optimization::kNone, G_a_power);
G_b.Spectrum(Aec3Optimization::kNone, G_b_power);
EXPECT_LT(std::accumulate(G_a_power.begin(), G_a_power.end(), 0.),
std::accumulate(G_b_power.begin(), G_b_power.end(), 0.));
}
}
// Verifies that the gain increases after an echo path change.
// TODO(peah): Correct and reactivate this test.
TEST(RefinedFilterUpdateGain, DISABLED_EchoPathChangeBehavior) {
for (size_t filter_length_blocks : {12, 20, 30}) {
SCOPED_TRACE(ProduceDebugText(filter_length_blocks));
std::vector<int> blocks_with_echo_path_changes;
std::vector<int> blocks_with_saturation;
blocks_with_echo_path_changes.push_back(99);
std::array<float, kBlockSize> e;
std::array<float, kBlockSize> y;
FftData G_a;
FftData G_b;
std::array<float, kFftLengthBy2Plus1> G_a_power;
std::array<float, kFftLengthBy2Plus1> G_b_power;
RunFilterUpdateTest(100, 65, filter_length_blocks,
blocks_with_echo_path_changes, blocks_with_saturation,
false, &e, &y, &G_a);
RunFilterUpdateTest(101, 65, filter_length_blocks,
blocks_with_echo_path_changes, blocks_with_saturation,
false, &e, &y, &G_b);
G_a.Spectrum(Aec3Optimization::kNone, G_a_power);
G_b.Spectrum(Aec3Optimization::kNone, G_b_power);
EXPECT_LT(std::accumulate(G_a_power.begin(), G_a_power.end(), 0.),
std::accumulate(G_b_power.begin(), G_b_power.end(), 0.));
}
}
} // namespace webrtc