Update filter analyzer for multi channel
Multi-channel behaviors introduced in this CL: - All filters are analyzed independently. The filtering is considered consistent if any filter is consistent. - The filter echo path gain used to detect saturation is maxed across capture channels. - The filter delay is taken to be the minimum of all filters: Any module that looks in the render data starting from the filter delay will iterate over all render audio present in any channel. - The FilterAnalyzer will consider a render block to be active if any render channel has activity. The changes in the CL has been shown to be bitexact on a large set of mono recordings. Bug: webrtc:10913 Change-Id: I1e360cd7136ee82d1f6e0f8a1459806e83f4426d Reviewed-on: https://webrtc-review.googlesource.com/c/src/+/155363 Reviewed-by: Per Åhgren <peah@webrtc.org> Commit-Queue: Sam Zackrisson <saza@webrtc.org> Cr-Commit-Position: refs/heads/master@{#29408}
This commit is contained in:
parent
43bd7601d7
commit
46b0140172
@ -346,11 +346,14 @@ TEST(AdaptiveFirFilter, FilterAndAdapt) {
|
|||||||
config.filter.main.length_blocks, config.filter.main.length_blocks,
|
config.filter.main.length_blocks, config.filter.main.length_blocks,
|
||||||
config.filter.config_change_duration_blocks, num_render_channels,
|
config.filter.config_change_duration_blocks, num_render_channels,
|
||||||
DetectOptimization(), &data_dumper);
|
DetectOptimization(), &data_dumper);
|
||||||
std::vector<std::array<float, kFftLengthBy2Plus1>> H2(
|
std::vector<std::vector<std::array<float, kFftLengthBy2Plus1>>> H2(
|
||||||
|
kNumCaptureChannels, std::vector<std::array<float, kFftLengthBy2Plus1>>(
|
||||||
filter.max_filter_size_partitions(),
|
filter.max_filter_size_partitions(),
|
||||||
std::array<float, kFftLengthBy2Plus1>());
|
std::array<float, kFftLengthBy2Plus1>()));
|
||||||
std::vector<float> h(
|
std::vector<std::vector<float>> h(
|
||||||
GetTimeDomainLength(filter.max_filter_size_partitions()), 0.f);
|
kNumCaptureChannels,
|
||||||
|
std::vector<float>(
|
||||||
|
GetTimeDomainLength(filter.max_filter_size_partitions()), 0.f));
|
||||||
Aec3Fft fft;
|
Aec3Fft fft;
|
||||||
config.delay.default_delay = 1;
|
config.delay.default_delay = 1;
|
||||||
std::unique_ptr<RenderDelayBuffer> render_delay_buffer(
|
std::unique_ptr<RenderDelayBuffer> render_delay_buffer(
|
||||||
@ -454,11 +457,11 @@ TEST(AdaptiveFirFilter, FilterAndAdapt) {
|
|||||||
render_buffer->SpectralSum(filter.SizePartitions(), &render_power);
|
render_buffer->SpectralSum(filter.SizePartitions(), &render_power);
|
||||||
gain.Compute(render_power, render_signal_analyzer, E,
|
gain.Compute(render_power, render_signal_analyzer, E,
|
||||||
filter.SizePartitions(), false, &G);
|
filter.SizePartitions(), false, &G);
|
||||||
filter.Adapt(*render_buffer, G, &h);
|
filter.Adapt(*render_buffer, G, &h[0]);
|
||||||
aec_state.HandleEchoPathChange(EchoPathVariability(
|
aec_state.HandleEchoPathChange(EchoPathVariability(
|
||||||
false, EchoPathVariability::DelayAdjustment::kNone, false));
|
false, EchoPathVariability::DelayAdjustment::kNone, false));
|
||||||
|
|
||||||
filter.ComputeFrequencyResponse(&H2);
|
filter.ComputeFrequencyResponse(&H2[0]);
|
||||||
aec_state.Update(delay_estimate, H2, h, *render_buffer, E2_main, Y2,
|
aec_state.Update(delay_estimate, H2, h, *render_buffer, E2_main, Y2,
|
||||||
output);
|
output);
|
||||||
}
|
}
|
||||||
|
|||||||
@ -66,18 +66,26 @@ AecState::AecState(const EchoCanceller3Config& config,
|
|||||||
filter_quality_state_(config_),
|
filter_quality_state_(config_),
|
||||||
erl_estimator_(2 * kNumBlocksPerSecond),
|
erl_estimator_(2 * kNumBlocksPerSecond),
|
||||||
erle_estimator_(2 * kNumBlocksPerSecond, config_, num_capture_channels),
|
erle_estimator_(2 * kNumBlocksPerSecond, config_, num_capture_channels),
|
||||||
filter_analyzer_(config_),
|
max_echo_path_gain_(config_.ep_strength.default_gain),
|
||||||
|
filter_analyzers_(num_capture_channels),
|
||||||
echo_audibility_(
|
echo_audibility_(
|
||||||
config_.echo_audibility.use_stationarity_properties_at_init),
|
config_.echo_audibility.use_stationarity_properties_at_init),
|
||||||
reverb_model_estimator_(config_),
|
reverb_model_estimator_(config_),
|
||||||
subtractor_output_analyzers_(num_capture_channels) {}
|
subtractor_output_analyzers_(num_capture_channels) {
|
||||||
|
for (size_t ch = 0; ch < num_capture_channels; ++ch) {
|
||||||
|
filter_analyzers_[ch] = std::make_unique<FilterAnalyzer>(config_);
|
||||||
|
}
|
||||||
|
}
|
||||||
|
|
||||||
AecState::~AecState() = default;
|
AecState::~AecState() = default;
|
||||||
|
|
||||||
void AecState::HandleEchoPathChange(
|
void AecState::HandleEchoPathChange(
|
||||||
const EchoPathVariability& echo_path_variability) {
|
const EchoPathVariability& echo_path_variability) {
|
||||||
const auto full_reset = [&]() {
|
const auto full_reset = [&]() {
|
||||||
filter_analyzer_.Reset();
|
for (auto& filter_analyzer : filter_analyzers_) {
|
||||||
|
filter_analyzer->Reset();
|
||||||
|
}
|
||||||
|
max_echo_path_gain_ = config_.ep_strength.default_gain;
|
||||||
capture_signal_saturation_ = false;
|
capture_signal_saturation_ = false;
|
||||||
strong_not_saturated_render_blocks_ = 0;
|
strong_not_saturated_render_blocks_ = 0;
|
||||||
blocks_with_active_render_ = 0;
|
blocks_with_active_render_ = 0;
|
||||||
@ -104,26 +112,43 @@ void AecState::HandleEchoPathChange(
|
|||||||
|
|
||||||
void AecState::Update(
|
void AecState::Update(
|
||||||
const absl::optional<DelayEstimate>& external_delay,
|
const absl::optional<DelayEstimate>& external_delay,
|
||||||
const std::vector<std::array<float, kFftLengthBy2Plus1>>&
|
rtc::ArrayView<const std::vector<std::array<float, kFftLengthBy2Plus1>>>
|
||||||
adaptive_filter_frequency_response,
|
adaptive_filter_frequency_response,
|
||||||
const std::vector<float>& adaptive_filter_impulse_response,
|
rtc::ArrayView<const std::vector<float>> adaptive_filter_impulse_response,
|
||||||
const RenderBuffer& render_buffer,
|
const RenderBuffer& render_buffer,
|
||||||
const std::array<float, kFftLengthBy2Plus1>& E2_main,
|
const std::array<float, kFftLengthBy2Plus1>& E2_main,
|
||||||
const std::array<float, kFftLengthBy2Plus1>& Y2,
|
const std::array<float, kFftLengthBy2Plus1>& Y2,
|
||||||
rtc::ArrayView<const SubtractorOutput> subtractor_output) {
|
rtc::ArrayView<const SubtractorOutput> subtractor_output) {
|
||||||
RTC_DCHECK_EQ(subtractor_output.size(), subtractor_output_analyzers_.size());
|
const size_t num_capture_channels = filter_analyzers_.size();
|
||||||
|
RTC_DCHECK_EQ(num_capture_channels, subtractor_output.size());
|
||||||
|
RTC_DCHECK_EQ(num_capture_channels, subtractor_output_analyzers_.size());
|
||||||
|
RTC_DCHECK_EQ(num_capture_channels,
|
||||||
|
adaptive_filter_frequency_response.size());
|
||||||
|
RTC_DCHECK_EQ(num_capture_channels, adaptive_filter_impulse_response.size());
|
||||||
|
|
||||||
// Analyze the filter output.
|
// Analyze the filter outputs and filters.
|
||||||
|
bool any_filter_converged = false;
|
||||||
|
bool all_filters_diverged = true;
|
||||||
|
bool any_filter_consistent = false;
|
||||||
|
max_echo_path_gain_ = 0.f;
|
||||||
for (size_t ch = 0; ch < subtractor_output.size(); ++ch) {
|
for (size_t ch = 0; ch < subtractor_output.size(); ++ch) {
|
||||||
subtractor_output_analyzers_[ch].Update(subtractor_output[ch]);
|
subtractor_output_analyzers_[ch].Update(subtractor_output[ch]);
|
||||||
}
|
any_filter_converged = any_filter_converged ||
|
||||||
|
subtractor_output_analyzers_[ch].ConvergedFilter();
|
||||||
|
all_filters_diverged = all_filters_diverged &&
|
||||||
|
subtractor_output_analyzers_[ch].DivergedFilter();
|
||||||
|
|
||||||
// Analyze the properties of the filter.
|
filter_analyzers_[ch]->Update(adaptive_filter_impulse_response[ch],
|
||||||
filter_analyzer_.Update(adaptive_filter_impulse_response, render_buffer);
|
render_buffer);
|
||||||
|
any_filter_consistent =
|
||||||
|
any_filter_consistent || filter_analyzers_[ch]->Consistent();
|
||||||
|
max_echo_path_gain_ =
|
||||||
|
std::max(max_echo_path_gain_, filter_analyzers_[ch]->Gain());
|
||||||
|
}
|
||||||
|
|
||||||
// Estimate the direct path delay of the filter.
|
// Estimate the direct path delay of the filter.
|
||||||
if (config_.filter.use_linear_filter) {
|
if (config_.filter.use_linear_filter) {
|
||||||
delay_state_.Update(filter_analyzer_, external_delay,
|
delay_state_.Update(filter_analyzers_, external_delay,
|
||||||
strong_not_saturated_render_blocks_);
|
strong_not_saturated_render_blocks_);
|
||||||
}
|
}
|
||||||
|
|
||||||
@ -170,7 +195,7 @@ void AecState::Update(
|
|||||||
/*channel=*/0);
|
/*channel=*/0);
|
||||||
const auto& X2_input_erle = X2_reverb;
|
const auto& X2_input_erle = X2_reverb;
|
||||||
|
|
||||||
erle_estimator_.Update(render_buffer, adaptive_filter_frequency_response,
|
erle_estimator_.Update(render_buffer, adaptive_filter_frequency_response[0],
|
||||||
X2_input_erle, Y2, E2_main,
|
X2_input_erle, Y2, E2_main,
|
||||||
subtractor_output_analyzers_[0].ConvergedFilter(),
|
subtractor_output_analyzers_[0].ConvergedFilter(),
|
||||||
config_.erle.onset_detection);
|
config_.erle.onset_detection);
|
||||||
@ -188,24 +213,22 @@ void AecState::Update(
|
|||||||
|
|
||||||
// Detect whether the transparent mode should be activated.
|
// Detect whether the transparent mode should be activated.
|
||||||
transparent_state_.Update(delay_state_.DirectPathFilterDelay(),
|
transparent_state_.Update(delay_state_.DirectPathFilterDelay(),
|
||||||
filter_analyzer_.Consistent(),
|
any_filter_consistent, any_filter_converged,
|
||||||
subtractor_output_analyzers_[0].ConvergedFilter(),
|
all_filters_diverged, active_render,
|
||||||
subtractor_output_analyzers_[0].DivergedFilter(),
|
SaturatedCapture());
|
||||||
active_render, SaturatedCapture());
|
|
||||||
|
|
||||||
// Analyze the quality of the filter.
|
// Analyze the quality of the filter.
|
||||||
filter_quality_state_.Update(
|
filter_quality_state_.Update(active_render, TransparentMode(),
|
||||||
active_render, TransparentMode(), SaturatedCapture(),
|
SaturatedCapture(), external_delay,
|
||||||
filter_analyzer_.Consistent(), external_delay,
|
any_filter_converged);
|
||||||
subtractor_output_analyzers_[0].ConvergedFilter());
|
|
||||||
|
|
||||||
// Update the reverb estimate.
|
// Update the reverb estimate.
|
||||||
const bool stationary_block =
|
const bool stationary_block =
|
||||||
config_.echo_audibility.use_stationarity_properties &&
|
config_.echo_audibility.use_stationarity_properties &&
|
||||||
echo_audibility_.IsBlockStationary();
|
echo_audibility_.IsBlockStationary();
|
||||||
|
|
||||||
reverb_model_estimator_.Update(filter_analyzer_.GetAdjustedFilter(),
|
reverb_model_estimator_.Update(filter_analyzers_[0]->GetAdjustedFilter(),
|
||||||
adaptive_filter_frequency_response,
|
adaptive_filter_frequency_response[0],
|
||||||
erle_estimator_.GetInstLinearQualityEstimate(),
|
erle_estimator_.GetInstLinearQualityEstimate(),
|
||||||
delay_state_.DirectPathFilterDelay(),
|
delay_state_.DirectPathFilterDelay(),
|
||||||
UsableLinearEstimate(), stationary_block);
|
UsableLinearEstimate(), stationary_block);
|
||||||
@ -217,18 +240,16 @@ void AecState::Update(
|
|||||||
data_dumper_->DumpRaw("aec3_erle", Erle()[0]);
|
data_dumper_->DumpRaw("aec3_erle", Erle()[0]);
|
||||||
data_dumper_->DumpRaw("aec3_usable_linear_estimate", UsableLinearEstimate());
|
data_dumper_->DumpRaw("aec3_usable_linear_estimate", UsableLinearEstimate());
|
||||||
data_dumper_->DumpRaw("aec3_transparent_mode", TransparentMode());
|
data_dumper_->DumpRaw("aec3_transparent_mode", TransparentMode());
|
||||||
data_dumper_->DumpRaw("aec3_filter_delay", filter_analyzer_.DelayBlocks());
|
data_dumper_->DumpRaw("aec3_filter_delay",
|
||||||
|
filter_analyzers_[0]->DelayBlocks());
|
||||||
|
|
||||||
data_dumper_->DumpRaw("aec3_consistent_filter",
|
data_dumper_->DumpRaw("aec3_any_filter_consistent", any_filter_consistent);
|
||||||
filter_analyzer_.Consistent());
|
|
||||||
data_dumper_->DumpRaw("aec3_initial_state",
|
data_dumper_->DumpRaw("aec3_initial_state",
|
||||||
initial_state_.InitialStateActive());
|
initial_state_.InitialStateActive());
|
||||||
data_dumper_->DumpRaw("aec3_capture_saturation", SaturatedCapture());
|
data_dumper_->DumpRaw("aec3_capture_saturation", SaturatedCapture());
|
||||||
data_dumper_->DumpRaw("aec3_echo_saturation", SaturatedEcho());
|
data_dumper_->DumpRaw("aec3_echo_saturation", SaturatedEcho());
|
||||||
data_dumper_->DumpRaw("aec3_converged_filter",
|
data_dumper_->DumpRaw("aec3_any_filter_converged", any_filter_converged);
|
||||||
subtractor_output_analyzers_[0].ConvergedFilter());
|
data_dumper_->DumpRaw("aec3_all_filters_diverged", all_filters_diverged);
|
||||||
data_dumper_->DumpRaw("aec3_diverged_filter",
|
|
||||||
subtractor_output_analyzers_[0].DivergedFilter());
|
|
||||||
|
|
||||||
data_dumper_->DumpRaw("aec3_external_delay_avaliable",
|
data_dumper_->DumpRaw("aec3_external_delay_avaliable",
|
||||||
external_delay ? 1 : 0);
|
external_delay ? 1 : 0);
|
||||||
@ -268,7 +289,7 @@ AecState::FilterDelay::FilterDelay(const EchoCanceller3Config& config)
|
|||||||
: delay_headroom_samples_(config.delay.delay_headroom_samples) {}
|
: delay_headroom_samples_(config.delay.delay_headroom_samples) {}
|
||||||
|
|
||||||
void AecState::FilterDelay::Update(
|
void AecState::FilterDelay::Update(
|
||||||
const FilterAnalyzer& filter_analyzer,
|
const std::vector<std::unique_ptr<FilterAnalyzer>>& filter_analyzers,
|
||||||
const absl::optional<DelayEstimate>& external_delay,
|
const absl::optional<DelayEstimate>& external_delay,
|
||||||
size_t blocks_with_proper_filter_adaptation) {
|
size_t blocks_with_proper_filter_adaptation) {
|
||||||
// Update the delay based on the external delay.
|
// Update the delay based on the external delay.
|
||||||
@ -285,7 +306,12 @@ void AecState::FilterDelay::Update(
|
|||||||
if (delay_estimator_may_not_have_converged && external_delay_) {
|
if (delay_estimator_may_not_have_converged && external_delay_) {
|
||||||
filter_delay_blocks_ = delay_headroom_samples_ / kBlockSize;
|
filter_delay_blocks_ = delay_headroom_samples_ / kBlockSize;
|
||||||
} else {
|
} else {
|
||||||
filter_delay_blocks_ = filter_analyzer.DelayBlocks();
|
// Conservatively use the min delay among the filters.
|
||||||
|
filter_delay_blocks_ = filter_analyzers[0]->DelayBlocks();
|
||||||
|
for (size_t ch = 1; ch < filter_analyzers.size(); ++ch) {
|
||||||
|
filter_delay_blocks_ =
|
||||||
|
std::min(filter_delay_blocks_, filter_analyzers[ch]->DelayBlocks());
|
||||||
|
}
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
|
||||||
@ -306,16 +332,16 @@ void AecState::TransparentMode::Reset() {
|
|||||||
}
|
}
|
||||||
|
|
||||||
void AecState::TransparentMode::Update(int filter_delay_blocks,
|
void AecState::TransparentMode::Update(int filter_delay_blocks,
|
||||||
bool consistent_filter,
|
bool any_filter_consistent,
|
||||||
bool converged_filter,
|
bool any_filter_converged,
|
||||||
bool diverged_filter,
|
bool all_filters_diverged,
|
||||||
bool active_render,
|
bool active_render,
|
||||||
bool saturated_capture) {
|
bool saturated_capture) {
|
||||||
++capture_block_counter_;
|
++capture_block_counter_;
|
||||||
strong_not_saturated_render_blocks_ +=
|
strong_not_saturated_render_blocks_ +=
|
||||||
active_render && !saturated_capture ? 1 : 0;
|
active_render && !saturated_capture ? 1 : 0;
|
||||||
|
|
||||||
if (consistent_filter && filter_delay_blocks < 5) {
|
if (any_filter_consistent && filter_delay_blocks < 5) {
|
||||||
sane_filter_observed_ = true;
|
sane_filter_observed_ = true;
|
||||||
active_blocks_since_sane_filter_ = 0;
|
active_blocks_since_sane_filter_ = 0;
|
||||||
} else if (active_render) {
|
} else if (active_render) {
|
||||||
@ -331,7 +357,7 @@ void AecState::TransparentMode::Update(int filter_delay_blocks,
|
|||||||
active_blocks_since_sane_filter_ <= 30 * kNumBlocksPerSecond;
|
active_blocks_since_sane_filter_ <= 30 * kNumBlocksPerSecond;
|
||||||
}
|
}
|
||||||
|
|
||||||
if (converged_filter) {
|
if (any_filter_converged) {
|
||||||
recent_convergence_during_activity_ = true;
|
recent_convergence_during_activity_ = true;
|
||||||
active_non_converged_sequence_size_ = 0;
|
active_non_converged_sequence_size_ = 0;
|
||||||
non_converged_sequence_size_ = 0;
|
non_converged_sequence_size_ = 0;
|
||||||
@ -347,7 +373,7 @@ void AecState::TransparentMode::Update(int filter_delay_blocks,
|
|||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
|
||||||
if (!diverged_filter) {
|
if (!all_filters_diverged) {
|
||||||
diverged_sequence_size_ = 0;
|
diverged_sequence_size_ = 0;
|
||||||
} else if (++diverged_sequence_size_ >= 60) {
|
} else if (++diverged_sequence_size_ >= 60) {
|
||||||
// TODO(peah): Change these lines to ensure proper triggering of usable
|
// TODO(peah): Change these lines to ensure proper triggering of usable
|
||||||
@ -387,16 +413,15 @@ void AecState::FilteringQualityAnalyzer::Update(
|
|||||||
bool active_render,
|
bool active_render,
|
||||||
bool transparent_mode,
|
bool transparent_mode,
|
||||||
bool saturated_capture,
|
bool saturated_capture,
|
||||||
bool consistent_estimate_,
|
|
||||||
const absl::optional<DelayEstimate>& external_delay,
|
const absl::optional<DelayEstimate>& external_delay,
|
||||||
bool converged_filter) {
|
bool any_filter_converged) {
|
||||||
// Update blocks counter.
|
// Update blocks counter.
|
||||||
const bool filter_update = active_render && !saturated_capture;
|
const bool filter_update = active_render && !saturated_capture;
|
||||||
filter_update_blocks_since_reset_ += filter_update ? 1 : 0;
|
filter_update_blocks_since_reset_ += filter_update ? 1 : 0;
|
||||||
filter_update_blocks_since_start_ += filter_update ? 1 : 0;
|
filter_update_blocks_since_start_ += filter_update ? 1 : 0;
|
||||||
|
|
||||||
// Store convergence flag when observed.
|
// Store convergence flag when observed.
|
||||||
convergence_seen_ = convergence_seen_ || converged_filter;
|
convergence_seen_ = convergence_seen_ || any_filter_converged;
|
||||||
|
|
||||||
// Verify requirements for achieving a decent filter. The requirements for
|
// Verify requirements for achieving a decent filter. The requirements for
|
||||||
// filter adaptation at call startup are more restrictive than after an
|
// filter adaptation at call startup are more restrictive than after an
|
||||||
|
|||||||
@ -57,7 +57,7 @@ class AecState {
|
|||||||
}
|
}
|
||||||
|
|
||||||
// Returns the estimated echo path gain.
|
// Returns the estimated echo path gain.
|
||||||
float EchoPathGain() const { return filter_analyzer_.Gain(); }
|
float EchoPathGain() const { return max_echo_path_gain_; }
|
||||||
|
|
||||||
// Returns whether the render signal is currently active.
|
// Returns whether the render signal is currently active.
|
||||||
bool ActiveRender() const { return blocks_with_active_render_ > 200; }
|
bool ActiveRender() const { return blocks_with_active_render_ > 200; }
|
||||||
@ -131,10 +131,11 @@ class AecState {
|
|||||||
// Updates the aec state.
|
// Updates the aec state.
|
||||||
// TODO(bugs.webrtc.org/10913): Handle multi-channel adaptive filter response.
|
// TODO(bugs.webrtc.org/10913): Handle multi-channel adaptive filter response.
|
||||||
// TODO(bugs.webrtc.org/10913): Compute multi-channel ERL, ERLE, and reverb.
|
// TODO(bugs.webrtc.org/10913): Compute multi-channel ERL, ERLE, and reverb.
|
||||||
void Update(const absl::optional<DelayEstimate>& external_delay,
|
void Update(
|
||||||
const std::vector<std::array<float, kFftLengthBy2Plus1>>&
|
const absl::optional<DelayEstimate>& external_delay,
|
||||||
|
rtc::ArrayView<const std::vector<std::array<float, kFftLengthBy2Plus1>>>
|
||||||
adaptive_filter_frequency_response,
|
adaptive_filter_frequency_response,
|
||||||
const std::vector<float>& adaptive_filter_impulse_response,
|
rtc::ArrayView<const std::vector<float>> adaptive_filter_impulse_response,
|
||||||
const RenderBuffer& render_buffer,
|
const RenderBuffer& render_buffer,
|
||||||
const std::array<float, kFftLengthBy2Plus1>& E2_main,
|
const std::array<float, kFftLengthBy2Plus1>& E2_main,
|
||||||
const std::array<float, kFftLengthBy2Plus1>& Y2,
|
const std::array<float, kFftLengthBy2Plus1>& Y2,
|
||||||
@ -142,7 +143,8 @@ class AecState {
|
|||||||
|
|
||||||
// Returns filter length in blocks.
|
// Returns filter length in blocks.
|
||||||
int FilterLengthBlocks() const {
|
int FilterLengthBlocks() const {
|
||||||
return filter_analyzer_.FilterLengthBlocks();
|
// All filters have the same length, so arbitrarily return channel 0 length.
|
||||||
|
return filter_analyzers_[/*channel=*/0]->FilterLengthBlocks();
|
||||||
}
|
}
|
||||||
|
|
||||||
private:
|
private:
|
||||||
@ -191,7 +193,8 @@ class AecState {
|
|||||||
int DirectPathFilterDelay() const { return filter_delay_blocks_; }
|
int DirectPathFilterDelay() const { return filter_delay_blocks_; }
|
||||||
|
|
||||||
// Updates the delay estimates based on new data.
|
// Updates the delay estimates based on new data.
|
||||||
void Update(const FilterAnalyzer& filter_analyzer,
|
void Update(
|
||||||
|
const std::vector<std::unique_ptr<FilterAnalyzer>>& filter_analyzer,
|
||||||
const absl::optional<DelayEstimate>& external_delay,
|
const absl::optional<DelayEstimate>& external_delay,
|
||||||
size_t blocks_with_proper_filter_adaptation);
|
size_t blocks_with_proper_filter_adaptation);
|
||||||
|
|
||||||
@ -216,9 +219,9 @@ class AecState {
|
|||||||
|
|
||||||
// Updates the detection deciscion based on new data.
|
// Updates the detection deciscion based on new data.
|
||||||
void Update(int filter_delay_blocks,
|
void Update(int filter_delay_blocks,
|
||||||
bool consistent_filter,
|
bool any_filter_consistent,
|
||||||
bool converged_filter,
|
bool any_filter_converged,
|
||||||
bool diverged_filter,
|
bool all_filters_diverged,
|
||||||
bool active_render,
|
bool active_render,
|
||||||
bool saturated_capture);
|
bool saturated_capture);
|
||||||
|
|
||||||
@ -257,9 +260,8 @@ class AecState {
|
|||||||
void Update(bool active_render,
|
void Update(bool active_render,
|
||||||
bool transparent_mode,
|
bool transparent_mode,
|
||||||
bool saturated_capture,
|
bool saturated_capture,
|
||||||
bool consistent_estimate_,
|
|
||||||
const absl::optional<DelayEstimate>& external_delay,
|
const absl::optional<DelayEstimate>& external_delay,
|
||||||
bool converged_filter);
|
bool any_filter_converged);
|
||||||
|
|
||||||
private:
|
private:
|
||||||
bool usable_linear_estimate_ = false;
|
bool usable_linear_estimate_ = false;
|
||||||
@ -290,8 +292,9 @@ class AecState {
|
|||||||
ErleEstimator erle_estimator_;
|
ErleEstimator erle_estimator_;
|
||||||
size_t strong_not_saturated_render_blocks_ = 0;
|
size_t strong_not_saturated_render_blocks_ = 0;
|
||||||
size_t blocks_with_active_render_ = 0;
|
size_t blocks_with_active_render_ = 0;
|
||||||
|
float max_echo_path_gain_;
|
||||||
bool capture_signal_saturation_ = false;
|
bool capture_signal_saturation_ = false;
|
||||||
FilterAnalyzer filter_analyzer_;
|
std::vector<std::unique_ptr<FilterAnalyzer>> filter_analyzers_;
|
||||||
absl::optional<DelayEstimate> external_delay_;
|
absl::optional<DelayEstimate> external_delay_;
|
||||||
EchoAudibility echo_audibility_;
|
EchoAudibility echo_audibility_;
|
||||||
ReverbModelEstimator reverb_model_estimator_;
|
ReverbModelEstimator reverb_model_estimator_;
|
||||||
|
|||||||
@ -55,17 +55,23 @@ void RunNormalUsageTest(size_t num_render_channels,
|
|||||||
y[ch].fill(1000.f);
|
y[ch].fill(1000.f);
|
||||||
}
|
}
|
||||||
Aec3Fft fft;
|
Aec3Fft fft;
|
||||||
std::vector<std::array<float, kFftLengthBy2Plus1>>
|
std::vector<std::vector<std::array<float, kFftLengthBy2Plus1>>>
|
||||||
converged_filter_frequency_response(10);
|
converged_filter_frequency_response(
|
||||||
for (auto& v : converged_filter_frequency_response) {
|
num_capture_channels,
|
||||||
|
std::vector<std::array<float, kFftLengthBy2Plus1>>(10));
|
||||||
|
for (auto& v_ch : converged_filter_frequency_response) {
|
||||||
|
for (auto& v : v_ch) {
|
||||||
v.fill(0.01f);
|
v.fill(0.01f);
|
||||||
}
|
}
|
||||||
std::vector<std::array<float, kFftLengthBy2Plus1>>
|
}
|
||||||
|
std::vector<std::vector<std::array<float, kFftLengthBy2Plus1>>>
|
||||||
diverged_filter_frequency_response = converged_filter_frequency_response;
|
diverged_filter_frequency_response = converged_filter_frequency_response;
|
||||||
converged_filter_frequency_response[2].fill(100.f);
|
converged_filter_frequency_response[0][2].fill(100.f);
|
||||||
converged_filter_frequency_response[2][0] = 1.f;
|
converged_filter_frequency_response[0][2][0] = 1.f;
|
||||||
std::vector<float> impulse_response(
|
std::vector<std::vector<float>> impulse_response(
|
||||||
GetTimeDomainLength(config.filter.main.length_blocks), 0.f);
|
num_capture_channels,
|
||||||
|
std::vector<float>(GetTimeDomainLength(config.filter.main.length_blocks),
|
||||||
|
0.f));
|
||||||
|
|
||||||
// Verify that linear AEC usability is true when the filter is converged
|
// Verify that linear AEC usability is true when the filter is converged
|
||||||
for (size_t band = 0; band < kNumBands; ++band) {
|
for (size_t band = 0; band < kNumBands; ++band) {
|
||||||
@ -243,20 +249,28 @@ TEST(AecState, ConvergedFilterDelay) {
|
|||||||
x.fill(0.f);
|
x.fill(0.f);
|
||||||
y.fill(0.f);
|
y.fill(0.f);
|
||||||
|
|
||||||
std::vector<std::array<float, kFftLengthBy2Plus1>> frequency_response(
|
std::vector<std::vector<std::array<float, kFftLengthBy2Plus1>>>
|
||||||
kFilterLengthBlocks);
|
frequency_response(
|
||||||
for (auto& v : frequency_response) {
|
kNumCaptureChannels,
|
||||||
|
std::vector<std::array<float, kFftLengthBy2Plus1>>(kFilterLengthBlocks));
|
||||||
|
for (auto& v_ch : frequency_response) {
|
||||||
|
for (auto& v : v_ch) {
|
||||||
v.fill(0.01f);
|
v.fill(0.01f);
|
||||||
}
|
}
|
||||||
|
}
|
||||||
|
|
||||||
std::vector<float> impulse_response(
|
std::vector<std::vector<float>> impulse_response(
|
||||||
GetTimeDomainLength(config.filter.main.length_blocks), 0.f);
|
kNumCaptureChannels,
|
||||||
|
std::vector<float>(GetTimeDomainLength(config.filter.main.length_blocks),
|
||||||
|
0.f));
|
||||||
|
|
||||||
// Verify that the filter delay for a converged filter is properly
|
// Verify that the filter delay for a converged filter is properly
|
||||||
// identified.
|
// identified.
|
||||||
for (int k = 0; k < kFilterLengthBlocks; ++k) {
|
for (int k = 0; k < kFilterLengthBlocks; ++k) {
|
||||||
std::fill(impulse_response.begin(), impulse_response.end(), 0.f);
|
for (auto& ir : impulse_response) {
|
||||||
impulse_response[k * kBlockSize + 1] = 1.f;
|
std::fill(ir.begin(), ir.end(), 0.f);
|
||||||
|
ir[k * kBlockSize + 1] = 1.f;
|
||||||
|
}
|
||||||
|
|
||||||
state.HandleEchoPathChange(echo_path_variability);
|
state.HandleEchoPathChange(echo_path_variability);
|
||||||
subtractor_output[0].ComputeMetrics(y);
|
subtractor_output[0].ComputeMetrics(y);
|
||||||
|
|||||||
@ -384,9 +384,9 @@ void EchoRemoverImpl::ProcessCapture(
|
|||||||
|
|
||||||
// Update the AEC state information.
|
// Update the AEC state information.
|
||||||
// TODO(bugs.webrtc.org/10913): Take all subtractors into account.
|
// TODO(bugs.webrtc.org/10913): Take all subtractors into account.
|
||||||
aec_state_.Update(external_delay, subtractor_.FilterFrequencyResponse()[0],
|
aec_state_.Update(external_delay, subtractor_.FilterFrequencyResponse(),
|
||||||
subtractor_.FilterImpulseResponse()[0], *render_buffer,
|
subtractor_.FilterImpulseResponse(), *render_buffer, E2[0],
|
||||||
E2[0], Y2[0], subtractor_output);
|
Y2[0], subtractor_output);
|
||||||
|
|
||||||
// Choose the linear output.
|
// Choose the linear output.
|
||||||
const auto& Y_fft = aec_state_.UseLinearFilterOutput() ? E : Y;
|
const auto& Y_fft = aec_state_.UseLinearFilterOutput() ? E : Y;
|
||||||
|
|||||||
@ -96,8 +96,8 @@ void FilterAnalyzer::AnalyzeRegion(
|
|||||||
filter_length_blocks_ = filter_time_domain.size() * (1.f / kBlockSize);
|
filter_length_blocks_ = filter_time_domain.size() * (1.f / kBlockSize);
|
||||||
|
|
||||||
consistent_estimate_ = consistent_filter_detector_.Detect(
|
consistent_estimate_ = consistent_filter_detector_.Detect(
|
||||||
h_highpass_, region_, render_buffer.Block(-delay_blocks_)[0][0],
|
h_highpass_, region_, render_buffer.Block(-delay_blocks_)[0], peak_index_,
|
||||||
peak_index_, delay_blocks_);
|
delay_blocks_);
|
||||||
}
|
}
|
||||||
|
|
||||||
void FilterAnalyzer::UpdateFilterGain(
|
void FilterAnalyzer::UpdateFilterGain(
|
||||||
@ -176,7 +176,7 @@ void FilterAnalyzer::ConsistentFilterDetector::Reset() {
|
|||||||
bool FilterAnalyzer::ConsistentFilterDetector::Detect(
|
bool FilterAnalyzer::ConsistentFilterDetector::Detect(
|
||||||
rtc::ArrayView<const float> filter_to_analyze,
|
rtc::ArrayView<const float> filter_to_analyze,
|
||||||
const FilterRegion& region,
|
const FilterRegion& region,
|
||||||
rtc::ArrayView<const float> x_block,
|
rtc::ArrayView<const std::vector<float>> x_block,
|
||||||
size_t peak_index,
|
size_t peak_index,
|
||||||
int delay_blocks) {
|
int delay_blocks) {
|
||||||
if (region.start_sample_ == 0) {
|
if (region.start_sample_ == 0) {
|
||||||
@ -212,9 +212,15 @@ bool FilterAnalyzer::ConsistentFilterDetector::Detect(
|
|||||||
}
|
}
|
||||||
|
|
||||||
if (significant_peak_) {
|
if (significant_peak_) {
|
||||||
const float x_energy = std::inner_product(x_block.begin(), x_block.end(),
|
bool active_render_block = false;
|
||||||
x_block.begin(), 0.f);
|
for (auto& x_channel : x_block) {
|
||||||
const bool active_render_block = x_energy > active_render_threshold_;
|
const float x_energy = std::inner_product(
|
||||||
|
x_channel.begin(), x_channel.end(), x_channel.begin(), 0.f);
|
||||||
|
if (x_energy > active_render_threshold_) {
|
||||||
|
active_render_block = true;
|
||||||
|
break;
|
||||||
|
}
|
||||||
|
}
|
||||||
|
|
||||||
if (consistent_delay_reference_ == delay_blocks) {
|
if (consistent_delay_reference_ == delay_blocks) {
|
||||||
if (active_render_block) {
|
if (active_render_block) {
|
||||||
|
|||||||
@ -33,6 +33,9 @@ class FilterAnalyzer {
|
|||||||
explicit FilterAnalyzer(const EchoCanceller3Config& config);
|
explicit FilterAnalyzer(const EchoCanceller3Config& config);
|
||||||
~FilterAnalyzer();
|
~FilterAnalyzer();
|
||||||
|
|
||||||
|
FilterAnalyzer(const FilterAnalyzer&) = delete;
|
||||||
|
FilterAnalyzer& operator=(const FilterAnalyzer&) = delete;
|
||||||
|
|
||||||
// Resets the analysis.
|
// Resets the analysis.
|
||||||
void Reset();
|
void Reset();
|
||||||
|
|
||||||
@ -82,7 +85,7 @@ class FilterAnalyzer {
|
|||||||
void Reset();
|
void Reset();
|
||||||
bool Detect(rtc::ArrayView<const float> filter_to_analyze,
|
bool Detect(rtc::ArrayView<const float> filter_to_analyze,
|
||||||
const FilterRegion& region,
|
const FilterRegion& region,
|
||||||
rtc::ArrayView<const float> x_block,
|
rtc::ArrayView<const std::vector<float>> x_block,
|
||||||
size_t peak_index,
|
size_t peak_index,
|
||||||
int delay_blocks);
|
int delay_blocks);
|
||||||
|
|
||||||
@ -110,8 +113,6 @@ class FilterAnalyzer {
|
|||||||
int filter_length_blocks_;
|
int filter_length_blocks_;
|
||||||
FilterRegion region_;
|
FilterRegion region_;
|
||||||
ConsistentFilterDetector consistent_filter_detector_;
|
ConsistentFilterDetector consistent_filter_detector_;
|
||||||
|
|
||||||
RTC_DISALLOW_COPY_AND_ASSIGN(FilterAnalyzer);
|
|
||||||
};
|
};
|
||||||
|
|
||||||
} // namespace webrtc
|
} // namespace webrtc
|
||||||
|
|||||||
@ -59,14 +59,19 @@ void RunFilterUpdateTest(int num_blocks_to_process,
|
|||||||
config.filter.shadow.length_blocks,
|
config.filter.shadow.length_blocks,
|
||||||
config.filter.config_change_duration_blocks,
|
config.filter.config_change_duration_blocks,
|
||||||
1, optimization, &data_dumper);
|
1, optimization, &data_dumper);
|
||||||
std::vector<std::array<float, kFftLengthBy2Plus1>> H2(
|
std::vector<std::vector<std::array<float, kFftLengthBy2Plus1>>> H2(
|
||||||
|
kNumChannels, std::vector<std::array<float, kFftLengthBy2Plus1>>(
|
||||||
main_filter.max_filter_size_partitions(),
|
main_filter.max_filter_size_partitions(),
|
||||||
std::array<float, kFftLengthBy2Plus1>());
|
std::array<float, kFftLengthBy2Plus1>()));
|
||||||
for (auto& H2_k : H2) {
|
for (auto& H2_ch : H2) {
|
||||||
|
for (auto& H2_k : H2_ch) {
|
||||||
H2_k.fill(0.f);
|
H2_k.fill(0.f);
|
||||||
}
|
}
|
||||||
std::vector<float> h(
|
}
|
||||||
GetTimeDomainLength(main_filter.max_filter_size_partitions()), 0.f);
|
std::vector<std::vector<float>> h(
|
||||||
|
kNumChannels,
|
||||||
|
std::vector<float>(
|
||||||
|
GetTimeDomainLength(main_filter.max_filter_size_partitions()), 0.f));
|
||||||
|
|
||||||
Aec3Fft fft;
|
Aec3Fft fft;
|
||||||
std::array<float, kBlockSize> x_old;
|
std::array<float, kBlockSize> x_old;
|
||||||
@ -183,15 +188,15 @@ void RunFilterUpdateTest(int num_blocks_to_process,
|
|||||||
main_filter.SizePartitions(), &render_power);
|
main_filter.SizePartitions(), &render_power);
|
||||||
|
|
||||||
std::array<float, kFftLengthBy2Plus1> erl;
|
std::array<float, kFftLengthBy2Plus1> erl;
|
||||||
ComputeErl(optimization, H2, erl);
|
ComputeErl(optimization, H2[0], erl);
|
||||||
main_gain.Compute(render_power, render_signal_analyzer, output[0], erl,
|
main_gain.Compute(render_power, render_signal_analyzer, output[0], erl,
|
||||||
main_filter.SizePartitions(), saturation, &G);
|
main_filter.SizePartitions(), saturation, &G);
|
||||||
main_filter.Adapt(*render_delay_buffer->GetRenderBuffer(), G, &h);
|
main_filter.Adapt(*render_delay_buffer->GetRenderBuffer(), G, &h[0]);
|
||||||
|
|
||||||
// Update the delay.
|
// Update the delay.
|
||||||
aec_state.HandleEchoPathChange(EchoPathVariability(
|
aec_state.HandleEchoPathChange(EchoPathVariability(
|
||||||
false, EchoPathVariability::DelayAdjustment::kNone, false));
|
false, EchoPathVariability::DelayAdjustment::kNone, false));
|
||||||
main_filter.ComputeFrequencyResponse(&H2);
|
main_filter.ComputeFrequencyResponse(&H2[0]);
|
||||||
aec_state.Update(delay_estimate, H2, h,
|
aec_state.Update(delay_estimate, H2, h,
|
||||||
*render_delay_buffer->GetRenderBuffer(), E2_main, Y2,
|
*render_delay_buffer->GetRenderBuffer(), E2_main, Y2,
|
||||||
output);
|
output);
|
||||||
|
|||||||
@ -28,7 +28,7 @@ TEST(ResidualEchoEstimator, BasicTest) {
|
|||||||
|
|
||||||
EchoCanceller3Config config;
|
EchoCanceller3Config config;
|
||||||
ResidualEchoEstimator estimator(config, num_render_channels);
|
ResidualEchoEstimator estimator(config, num_render_channels);
|
||||||
AecState aec_state(config, num_render_channels);
|
AecState aec_state(config, num_capture_channels);
|
||||||
std::unique_ptr<RenderDelayBuffer> render_delay_buffer(
|
std::unique_ptr<RenderDelayBuffer> render_delay_buffer(
|
||||||
RenderDelayBuffer::Create(config, kSampleRateHz,
|
RenderDelayBuffer::Create(config, kSampleRateHz,
|
||||||
num_render_channels));
|
num_render_channels));
|
||||||
@ -44,20 +44,26 @@ TEST(ResidualEchoEstimator, BasicTest) {
|
|||||||
kNumBands,
|
kNumBands,
|
||||||
std::vector<std::vector<float>>(num_render_channels,
|
std::vector<std::vector<float>>(num_render_channels,
|
||||||
std::vector<float>(kBlockSize, 0.f)));
|
std::vector<float>(kBlockSize, 0.f)));
|
||||||
std::vector<std::array<float, kFftLengthBy2Plus1>> H2(10);
|
std::vector<std::vector<std::array<float, kFftLengthBy2Plus1>>> H2(
|
||||||
|
num_capture_channels,
|
||||||
|
std::vector<std::array<float, kFftLengthBy2Plus1>>(10));
|
||||||
Random random_generator(42U);
|
Random random_generator(42U);
|
||||||
std::vector<SubtractorOutput> output(num_render_channels);
|
std::vector<SubtractorOutput> output(num_capture_channels);
|
||||||
std::array<float, kBlockSize> y;
|
std::array<float, kBlockSize> y;
|
||||||
absl::optional<DelayEstimate> delay_estimate;
|
absl::optional<DelayEstimate> delay_estimate;
|
||||||
|
|
||||||
for (auto& H2_k : H2) {
|
for (auto& H2_ch : H2) {
|
||||||
|
for (auto& H2_k : H2_ch) {
|
||||||
H2_k.fill(0.01f);
|
H2_k.fill(0.01f);
|
||||||
}
|
}
|
||||||
H2[2].fill(10.f);
|
H2_ch[2].fill(10.f);
|
||||||
H2[2][0] = 0.1f;
|
H2_ch[2][0] = 0.1f;
|
||||||
|
}
|
||||||
|
|
||||||
std::vector<float> h(
|
std::vector<std::vector<float>> h(
|
||||||
GetTimeDomainLength(config.filter.main.length_blocks), 0.f);
|
num_capture_channels,
|
||||||
|
std::vector<float>(
|
||||||
|
GetTimeDomainLength(config.filter.main.length_blocks), 0.f));
|
||||||
|
|
||||||
for (auto& subtractor_output : output) {
|
for (auto& subtractor_output : output) {
|
||||||
subtractor_output.Reset();
|
subtractor_output.Reset();
|
||||||
|
|||||||
@ -145,8 +145,8 @@ std::vector<float> RunSubtractorTest(
|
|||||||
|
|
||||||
aec_state.HandleEchoPathChange(EchoPathVariability(
|
aec_state.HandleEchoPathChange(EchoPathVariability(
|
||||||
false, EchoPathVariability::DelayAdjustment::kNone, false));
|
false, EchoPathVariability::DelayAdjustment::kNone, false));
|
||||||
aec_state.Update(delay_estimate, subtractor.FilterFrequencyResponse()[0],
|
aec_state.Update(delay_estimate, subtractor.FilterFrequencyResponse(),
|
||||||
subtractor.FilterImpulseResponse()[0],
|
subtractor.FilterImpulseResponse(),
|
||||||
*render_delay_buffer->GetRenderBuffer(), E2_main, Y2,
|
*render_delay_buffer->GetRenderBuffer(), E2_main, Y2,
|
||||||
output);
|
output);
|
||||||
}
|
}
|
||||||
|
|||||||
@ -97,14 +97,14 @@ TEST(SuppressionGain, BasicGainComputation) {
|
|||||||
|
|
||||||
// Ensure that the gain is no longer forced to zero.
|
// Ensure that the gain is no longer forced to zero.
|
||||||
for (int k = 0; k <= kNumBlocksPerSecond / 5 + 1; ++k) {
|
for (int k = 0; k <= kNumBlocksPerSecond / 5 + 1; ++k) {
|
||||||
aec_state.Update(delay_estimate, subtractor.FilterFrequencyResponse()[0],
|
aec_state.Update(delay_estimate, subtractor.FilterFrequencyResponse(),
|
||||||
subtractor.FilterImpulseResponse()[0],
|
subtractor.FilterImpulseResponse(),
|
||||||
*render_delay_buffer->GetRenderBuffer(), E2, Y2, output);
|
*render_delay_buffer->GetRenderBuffer(), E2, Y2, output);
|
||||||
}
|
}
|
||||||
|
|
||||||
for (int k = 0; k < 100; ++k) {
|
for (int k = 0; k < 100; ++k) {
|
||||||
aec_state.Update(delay_estimate, subtractor.FilterFrequencyResponse()[0],
|
aec_state.Update(delay_estimate, subtractor.FilterFrequencyResponse(),
|
||||||
subtractor.FilterImpulseResponse()[0],
|
subtractor.FilterImpulseResponse(),
|
||||||
*render_delay_buffer->GetRenderBuffer(), E2, Y2, output);
|
*render_delay_buffer->GetRenderBuffer(), E2, Y2, output);
|
||||||
suppression_gain.GetGain(E2, S2, R2, N2, analyzer, aec_state, x,
|
suppression_gain.GetGain(E2, S2, R2, N2, analyzer, aec_state, x,
|
||||||
&high_bands_gain, &g);
|
&high_bands_gain, &g);
|
||||||
@ -120,8 +120,8 @@ TEST(SuppressionGain, BasicGainComputation) {
|
|||||||
N2.fill(0.f);
|
N2.fill(0.f);
|
||||||
|
|
||||||
for (int k = 0; k < 100; ++k) {
|
for (int k = 0; k < 100; ++k) {
|
||||||
aec_state.Update(delay_estimate, subtractor.FilterFrequencyResponse()[0],
|
aec_state.Update(delay_estimate, subtractor.FilterFrequencyResponse(),
|
||||||
subtractor.FilterImpulseResponse()[0],
|
subtractor.FilterImpulseResponse(),
|
||||||
*render_delay_buffer->GetRenderBuffer(), E2, Y2, output);
|
*render_delay_buffer->GetRenderBuffer(), E2, Y2, output);
|
||||||
suppression_gain.GetGain(E2, S2, R2, N2, analyzer, aec_state, x,
|
suppression_gain.GetGain(E2, S2, R2, N2, analyzer, aec_state, x,
|
||||||
&high_bands_gain, &g);
|
&high_bands_gain, &g);
|
||||||
|
|||||||
Loading…
x
Reference in New Issue
Block a user