modules/audio_processing/aec3/suppression_gain.cc

/*
 *  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/suppression_gain.h"

#include <math.h>
#include <algorithm>
#include <functional>
#include <numeric>

#include "modules/audio_processing/aec3/moving_average.h"
#include "modules/audio_processing/aec3/vector_math.h"
#include "modules/audio_processing/logging/apm_data_dumper.h"
#include "rtc_base/atomicops.h"
#include "rtc_base/checks.h"
#include "system_wrappers/include/field_trial.h"

namespace webrtc {
namespace {

bool EnableNewSuppression() {
  return !field_trial::IsEnabled("WebRTC-Aec3NewSuppressionKillSwitch");
}

// Adjust the gains according to the presence of known external filters.
void AdjustForExternalFilters(std::array<float, kFftLengthBy2Plus1>* gain) {
  // Limit the low frequency gains to avoid the impact of the high-pass filter
  // on the lower-frequency gain influencing the overall achieved gain.
  (*gain)[0] = (*gain)[1] = std::min((*gain)[1], (*gain)[2]);

  // Limit the high frequency gains to avoid the impact of the anti-aliasing
  // filter on the upper-frequency gains influencing the overall achieved
  // gain. TODO(peah): Update this when new anti-aliasing filters are
  // implemented.
  constexpr size_t kAntiAliasingImpactLimit = (64 * 2000) / 8000;
  const float min_upper_gain = (*gain)[kAntiAliasingImpactLimit];
  std::for_each(
      gain->begin() + kAntiAliasingImpactLimit, gain->end() - 1,
      [min_upper_gain](float& a) { a = std::min(a, min_upper_gain); });
  (*gain)[kFftLengthBy2] = (*gain)[kFftLengthBy2Minus1];
}

// Computes the gain to apply for the bands beyond the first band.
float UpperBandsGain(
    const absl::optional<int>& narrow_peak_band,
    bool saturated_echo,
    const std::vector<std::vector<float>>& render,
    const std::array<float, kFftLengthBy2Plus1>& low_band_gain) {
  RTC_DCHECK_LT(0, render.size());
  if (render.size() == 1) {
    return 1.f;
  }

  if (narrow_peak_band &&
      (*narrow_peak_band > static_cast<int>(kFftLengthBy2Plus1 - 10))) {
    return 0.001f;
  }

  constexpr size_t kLowBandGainLimit = kFftLengthBy2 / 2;
  const float gain_below_8_khz = *std::min_element(
      low_band_gain.begin() + kLowBandGainLimit, low_band_gain.end());

  // Always attenuate the upper bands when there is saturated echo.
  if (saturated_echo) {
    return std::min(0.001f, gain_below_8_khz);
  }

  // Compute the upper and lower band energies.
  const auto sum_of_squares = [](float a, float b) { return a + b * b; };
  const float low_band_energy =
      std::accumulate(render[0].begin(), render[0].end(), 0.f, sum_of_squares);
  float high_band_energy = 0.f;
  for (size_t k = 1; k < render.size(); ++k) {
    const float energy = std::accumulate(render[k].begin(), render[k].end(),
                                         0.f, sum_of_squares);
    high_band_energy = std::max(high_band_energy, energy);
  }

  // If there is more power in the lower frequencies than the upper frequencies,
  // or if the power in upper frequencies is low, do not bound the gain in the
  // upper bands.
  float anti_howling_gain;
  constexpr float kThreshold = kBlockSize * 10.f * 10.f / 4.f;
  if (high_band_energy < std::max(low_band_energy, kThreshold)) {
    anti_howling_gain = 1.f;
  } else {
    // In all other cases, bound the gain for upper frequencies.
    RTC_DCHECK_LE(low_band_energy, high_band_energy);
    RTC_DCHECK_NE(0.f, high_band_energy);
    anti_howling_gain = 0.01f * sqrtf(low_band_energy / high_band_energy);
  }

  // Choose the gain as the minimum of the lower and upper gains.
  return std::min(gain_below_8_khz, anti_howling_gain);
}

// Scales the echo according to assessed audibility at the other end.
void WeightEchoForAudibility(const EchoCanceller3Config& config,
                             rtc::ArrayView<const float> echo,
                             rtc::ArrayView<float> weighted_echo,
                             rtc::ArrayView<float> one_by_weighted_echo) {
  RTC_DCHECK_EQ(kFftLengthBy2Plus1, echo.size());
  RTC_DCHECK_EQ(kFftLengthBy2Plus1, weighted_echo.size());
  RTC_DCHECK_EQ(kFftLengthBy2Plus1, one_by_weighted_echo.size());

  auto weigh = [](float threshold, float normalizer, size_t begin, size_t end,
                  rtc::ArrayView<const float> echo,
                  rtc::ArrayView<float> weighted_echo,
                  rtc::ArrayView<float> one_by_weighted_echo) {
    for (size_t k = begin; k < end; ++k) {
      if (echo[k] < threshold) {
        float tmp = (threshold - echo[k]) * normalizer;
        weighted_echo[k] = echo[k] * std::max(0.f, 1.f - tmp * tmp);
      } else {
        weighted_echo[k] = echo[k];
      }
      one_by_weighted_echo[k] =
          weighted_echo[k] > 0.f ? 1.f / weighted_echo[k] : 1.f;
    }
  };

  float threshold = config.echo_audibility.floor_power *
                    config.echo_audibility.audibility_threshold_lf;
  float normalizer = 1.f / (threshold - config.echo_audibility.floor_power);
  weigh(threshold, normalizer, 0, 3, echo, weighted_echo, one_by_weighted_echo);

  threshold = config.echo_audibility.floor_power *
              config.echo_audibility.audibility_threshold_mf;
  normalizer = 1.f / (threshold - config.echo_audibility.floor_power);
  weigh(threshold, normalizer, 3, 7, echo, weighted_echo, one_by_weighted_echo);

  threshold = config.echo_audibility.floor_power *
              config.echo_audibility.audibility_threshold_hf;
  normalizer = 1.f / (threshold - config.echo_audibility.floor_power);
  weigh(threshold, normalizer, 7, kFftLengthBy2Plus1, echo, weighted_echo,
        one_by_weighted_echo);
}

// Computes the gain to reduce the echo to a non audible level.
void GainToNoAudibleEchoFallback(
    const EchoCanceller3Config& config,
    bool low_noise_render,
    bool saturated_echo,
    bool linear_echo_estimate,
    const std::array<float, kFftLengthBy2Plus1>& nearend,
    const std::array<float, kFftLengthBy2Plus1>& weighted_echo,
    const std::array<float, kFftLengthBy2Plus1>& masker,
    const std::array<float, kFftLengthBy2Plus1>& min_gain,
    const std::array<float, kFftLengthBy2Plus1>& max_gain,
    const std::array<float, kFftLengthBy2Plus1>& one_by_weighted_echo,
    std::array<float, kFftLengthBy2Plus1>* gain) {
  float nearend_masking_margin = 0.f;
  if (linear_echo_estimate) {
    nearend_masking_margin =
        low_noise_render
            ? config.gain_mask.m9
            : (saturated_echo ? config.gain_mask.m2 : config.gain_mask.m3);
  } else {
    nearend_masking_margin = config.gain_mask.m7;
  }

  RTC_DCHECK_LE(0.f, nearend_masking_margin);
  RTC_DCHECK_GT(1.f, nearend_masking_margin);

  const float masker_margin =
      linear_echo_estimate ? config.gain_mask.m0 : config.gain_mask.m8;

  for (size_t k = 0; k < gain->size(); ++k) {
    // TODO(devicentepena): Experiment by removing the reverberation estimation
    // from the nearend signal before computing the gains.
    const float unity_gain_masker = std::max(nearend[k], masker[k]);
    RTC_DCHECK_LE(0.f, nearend_masking_margin * unity_gain_masker);
    if (weighted_echo[k] <= nearend_masking_margin * unity_gain_masker ||
        unity_gain_masker <= 0.f) {
      (*gain)[k] = 1.f;
    } else {
      RTC_DCHECK_LT(0.f, unity_gain_masker);
      (*gain)[k] =
          std::max(0.f, (1.f - config.gain_mask.gain_curve_slope *
                                   weighted_echo[k] / unity_gain_masker) *
                            config.gain_mask.gain_curve_offset);
      (*gain)[k] = std::max(masker_margin * masker[k] * one_by_weighted_echo[k],
                            (*gain)[k]);
    }

    (*gain)[k] = std::min(std::max((*gain)[k], min_gain[k]), max_gain[k]);
  }
}

// TODO(peah): Make adaptive to take the actual filter error into account.
constexpr size_t kUpperAccurateBandPlus1 = 29;

// Limits the gain in the frequencies for which the adaptive filter has not
// converged. Currently, these frequencies are not hardcoded to the frequencies
// which are typically not excited by speech.
// TODO(peah): Make adaptive to take the actual filter error into account.
void AdjustNonConvergedFrequencies(
    std::array<float, kFftLengthBy2Plus1>* gain) {
  constexpr float oneByBandsInSum =
      1 / static_cast<float>(kUpperAccurateBandPlus1 - 20);
  const float hf_gain_bound =
      std::accumulate(gain->begin() + 20,
                      gain->begin() + kUpperAccurateBandPlus1, 0.f) *
      oneByBandsInSum;

  std::for_each(gain->begin() + kUpperAccurateBandPlus1, gain->end(),
                [hf_gain_bound](float& a) { a = std::min(a, hf_gain_bound); });
}

}  // namespace

int SuppressionGain::instance_count_ = 0;

// Computes the gain to reduce the echo to a non audible level.
void SuppressionGain::GainToNoAudibleEcho(
    const std::array<float, kFftLengthBy2Plus1>& nearend,
    const std::array<float, kFftLengthBy2Plus1>& echo,
    const std::array<float, kFftLengthBy2Plus1>& masker,
    const std::array<float, kFftLengthBy2Plus1>& min_gain,
    const std::array<float, kFftLengthBy2Plus1>& max_gain,
    std::array<float, kFftLengthBy2Plus1>* gain) const {
  for (size_t k = 0; k < gain->size(); ++k) {
    float enr = echo[k] / (nearend[k] + 1.f);  // Echo-to-nearend ratio.
    float emr = echo[k] / (masker[k] + 1.f);   // Echo-to-masker (noise) ratio.
    float g = 1.0f;
    if (enr > enr_transparent_[k] && emr > emr_transparent_[k]) {
      g = (enr_suppress_[k] - enr) / (enr_suppress_[k] - enr_transparent_[k]);
      g = std::max(g, emr_transparent_[k] / emr);
    }
    (*gain)[k] = std::max(std::min(g, max_gain[k]), min_gain[k]);
  }
}

// TODO(peah): Add further optimizations, in particular for the divisions.
void SuppressionGain::LowerBandGain(
    bool low_noise_render,
    const AecState& aec_state,
    const std::array<float, kFftLengthBy2Plus1>& nearend,
    const std::array<float, kFftLengthBy2Plus1>& echo,
    const std::array<float, kFftLengthBy2Plus1>& comfort_noise,
    std::array<float, kFftLengthBy2Plus1>* gain) {
  const bool saturated_echo = aec_state.SaturatedEcho();
  const bool linear_echo_estimate = aec_state.UsableLinearEstimate();

  // Weight echo power in terms of audibility. // Precompute 1/weighted echo
  // (note that when the echo is zero, the precomputed value is never used).
  std::array<float, kFftLengthBy2Plus1> weighted_echo;
  std::array<float, kFftLengthBy2Plus1> one_by_weighted_echo;
  WeightEchoForAudibility(config_, echo, weighted_echo, one_by_weighted_echo);

  // Compute the minimum gain as the attenuating gain to put the signal just
  // above the zero sample values.
  std::array<float, kFftLengthBy2Plus1> min_gain;
  const float min_echo_power =
      low_noise_render ? config_.echo_audibility.low_render_limit
                       : config_.echo_audibility.normal_render_limit;
  if (!saturated_echo) {
    for (size_t k = 0; k < nearend.size(); ++k) {
      const float denom = std::min(nearend[k], weighted_echo[k]);
      min_gain[k] = denom > 0.f ? min_echo_power / denom : 1.f;
      min_gain[k] = std::min(min_gain[k], 1.f);
    }
    for (size_t k = 0; k < 6; ++k) {
      // Make sure the gains of the low frequencies do not decrease too
      // quickly after strong nearend.
      if (last_nearend_[k] > last_echo_[k]) {
        min_gain[k] =
            std::max(min_gain[k],
                     last_gain_[k] * config_.gain_updates.max_dec_factor_lf);
        min_gain[k] = std::min(min_gain[k], 1.f);
      }
    }
  } else {
    min_gain.fill(0.f);
  }

  // Compute the maximum gain by limiting the gain increase from the previous
  // gain.
  std::array<float, kFftLengthBy2Plus1> max_gain;
  for (size_t k = 0; k < gain->size(); ++k) {
    max_gain[k] =
        std::min(std::max(last_gain_[k] * config_.gain_updates.max_inc_factor,
                          config_.gain_updates.floor_first_increase),
                 1.f);
  }

  // Iteratively compute the gain required to attenuate the echo to a non
  // noticeable level.
  std::array<float, kFftLengthBy2Plus1> masker;
  if (enable_new_suppression_) {
    GainToNoAudibleEcho(nearend, weighted_echo, comfort_noise, min_gain,
                        max_gain, gain);
    AdjustForExternalFilters(gain);
  } else {
    gain->fill(0.f);
    for (int k = 0; k < 2; ++k) {
      std::copy(comfort_noise.begin(), comfort_noise.end(), masker.begin());
      GainToNoAudibleEchoFallback(config_, low_noise_render, saturated_echo,
                                  linear_echo_estimate, nearend, weighted_echo,
                                  masker, min_gain, max_gain,
                                  one_by_weighted_echo, gain);
      AdjustForExternalFilters(gain);
    }
  }

  // Adjust the gain for frequencies which have not yet converged.
  AdjustNonConvergedFrequencies(gain);

  // Store data required for the gain computation of the next block.
  std::copy(nearend.begin(), nearend.end(), last_nearend_.begin());
  std::copy(weighted_echo.begin(), weighted_echo.end(), last_echo_.begin());
  std::copy(gain->begin(), gain->end(), last_gain_.begin());
  aec3::VectorMath(optimization_).Sqrt(*gain);

  // Debug outputs for the purpose of development and analysis.
  data_dumper_->DumpRaw("aec3_suppressor_min_gain", min_gain);
  data_dumper_->DumpRaw("aec3_suppressor_max_gain", max_gain);
  data_dumper_->DumpRaw("aec3_suppressor_masker", masker);
}

SuppressionGain::SuppressionGain(const EchoCanceller3Config& config,
                                 Aec3Optimization optimization,
                                 int sample_rate_hz)
    : data_dumper_(
          new ApmDataDumper(rtc::AtomicOps::Increment(&instance_count_))),
      optimization_(optimization),
      config_(config),
      state_change_duration_blocks_(
          static_cast<int>(config_.filter.config_change_duration_blocks)),
      enable_new_suppression_(EnableNewSuppression()),
      moving_average_(kFftLengthBy2Plus1,
                      config.suppressor.nearend_average_blocks) {
  RTC_DCHECK_LT(0, state_change_duration_blocks_);
  one_by_state_change_duration_blocks_ = 1.f / state_change_duration_blocks_;
  last_gain_.fill(1.f);
  last_nearend_.fill(0.f);
  last_echo_.fill(0.f);

  // Compute per-band masking thresholds.
  constexpr size_t kLastLfBand = 5;
  constexpr size_t kFirstHfBand = 8;
  RTC_DCHECK_LT(kLastLfBand, kFirstHfBand);
  auto& lf = config.suppressor.mask_lf;
  auto& hf = config.suppressor.mask_hf;
  RTC_DCHECK_LT(lf.enr_transparent, lf.enr_suppress);
  RTC_DCHECK_LT(hf.enr_transparent, hf.enr_suppress);
  for (size_t k = 0; k < kFftLengthBy2Plus1; k++) {
    float a;
    if (k <= kLastLfBand) {
      a = 0.f;
    } else if (k < kFirstHfBand) {
      a = (k - kLastLfBand) / static_cast<float>(kFirstHfBand - kLastLfBand);
    } else {
      a = 1.f;
    }
    enr_transparent_[k] = (1 - a) * lf.enr_transparent + a * hf.enr_transparent;
    enr_suppress_[k] = (1 - a) * lf.enr_suppress + a * hf.enr_suppress;
    emr_transparent_[k] = (1 - a) * lf.emr_transparent + a * hf.emr_transparent;
  }
}

SuppressionGain::~SuppressionGain() = default;

void SuppressionGain::GetGain(
    const std::array<float, kFftLengthBy2Plus1>& nearend_spectrum,
    const std::array<float, kFftLengthBy2Plus1>& echo_spectrum,
    const std::array<float, kFftLengthBy2Plus1>& comfort_noise_spectrum,
    const FftData& linear_aec_fft,
    const FftData& capture_fft,
    const RenderSignalAnalyzer& render_signal_analyzer,
    const AecState& aec_state,
    const std::vector<std::vector<float>>& render,
    float* high_bands_gain,
    std::array<float, kFftLengthBy2Plus1>* low_band_gain) {
  RTC_DCHECK(high_bands_gain);
  RTC_DCHECK(low_band_gain);
  const auto& cfg = config_.suppressor;

  if (cfg.enforce_transparent) {
    low_band_gain->fill(1.f);
    *high_bands_gain = cfg.enforce_empty_higher_bands ? 0.f : 1.f;
    return;
  }

  std::array<float, kFftLengthBy2Plus1> nearend_average;
  moving_average_.Average(nearend_spectrum, nearend_average);

  // Compute gain for the lower band.
  bool low_noise_render = low_render_detector_.Detect(render);
  const absl::optional<int> narrow_peak_band =
      render_signal_analyzer.NarrowPeakBand();
  LowerBandGain(low_noise_render, aec_state, nearend_average, echo_spectrum,
                comfort_noise_spectrum, low_band_gain);

  // Limit the gain of the lower bands during start up and after resets.
  const float gain_upper_bound = aec_state.SuppressionGainLimit();
  if (gain_upper_bound < 1.f) {
    for (size_t k = 0; k < low_band_gain->size(); ++k) {
      (*low_band_gain)[k] = std::min((*low_band_gain)[k], gain_upper_bound);
    }
  }

  // Compute the gain for the upper bands.
  *high_bands_gain = UpperBandsGain(narrow_peak_band, aec_state.SaturatedEcho(),
                                    render, *low_band_gain);
  if (cfg.enforce_empty_higher_bands) {
    *high_bands_gain = 0.f;
  }
}

void SuppressionGain::SetInitialState(bool state) {
  initial_state_ = state;
  if (state) {
    initial_state_change_counter_ = state_change_duration_blocks_;
  } else {
    initial_state_change_counter_ = 0;
  }
}

// Detects when the render signal can be considered to have low power and
// consist of stationary noise.
bool SuppressionGain::LowNoiseRenderDetector::Detect(
    const std::vector<std::vector<float>>& render) {
  float x2_sum = 0.f;
  float x2_max = 0.f;
  for (auto x_k : render[0]) {
    const float x2 = x_k * x_k;
    x2_sum += x2;
    x2_max = std::max(x2_max, x2);
  }

  constexpr float kThreshold = 50.f * 50.f * 64.f;
  const bool low_noise_render =
      average_power_ < kThreshold && x2_max < 3 * average_power_;
  average_power_ = average_power_ * 0.9f + x2_sum * 0.1f;
  return low_noise_render;
}

}  // namespace webrtc