Finalization of the first version of EchoCanceller 3

webrtc/modules/audio_processing/aec3/suppression_filter.cc

Index: webrtc/modules/audio_processing/aec3/suppression_filter.cc
diff --git a/webrtc/modules/audio_processing/aec3/suppression_filter.cc b/webrtc/modules/audio_processing/aec3/suppression_filter.cc
new file mode 100644
index 0000000000000000000000000000000000000000..6eef3546da590172016f25ed771e1960656dadfa
--- /dev/null
+++ b/webrtc/modules/audio_processing/aec3/suppression_filter.cc
@@ -0,0 +1,180 @@
+/*
+ *  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 "webrtc/modules/audio_processing/aec3/suppression_filter.h"
+
+#include <math.h>
+#include <algorithm>
+#include <cstring>
+#include <functional>
+#include <numeric>
+
+#include "webrtc/modules/audio_processing/utility/ooura_fft.h"
+
+namespace webrtc {
+namespace {
+
+// Matlab code to produce table:
+// win = sqrt(hanning(63)); win = [0 ; win(1:32)];
+// fprintf(1, '\t%.14f, %.14f, %.14f,\n', win);

[aleloi, 2017/02/22 16:20:31]

This seems to be a few more than 32 numbers. It also goes from 0 to 1 and back
to 0, while [0 ; win(1:32)] goes from 0 to 1, I think. At least if Octave has
the same 'hanning()' as Matlab does. Maybe it's hanning(128)?

[peah-webrtc, 2017/02/22 23:51:40]

True. Good find!
Done.

+const float kSqrtHanning[kFftLength] = {
+    0.00000000000000f, 0.02454122852291f, 0.04906767432742f, 0.07356456359967f,
+    0.09801714032956f, 0.12241067519922f, 0.14673047445536f, 0.17096188876030f,
+    0.19509032201613f, 0.21910124015687f, 0.24298017990326f, 0.26671275747490f,
+    0.29028467725446f, 0.31368174039889f, 0.33688985339222f, 0.35989503653499f,
+    0.38268343236509f, 0.40524131400499f, 0.42755509343028f, 0.44961132965461f,
+    0.47139673682600f, 0.49289819222978f, 0.51410274419322f, 0.53499761988710f,
+    0.55557023301960f, 0.57580819141785f, 0.59569930449243f, 0.61523159058063f,
+    0.63439328416365f, 0.65317284295378f, 0.67155895484702f, 0.68954054473707f,
+    0.70710678118655f, 0.72424708295147f, 0.74095112535496f, 0.75720884650648f,
+    0.77301045336274f, 0.78834642762661f, 0.80320753148064f, 0.81758481315158f,
+    0.83146961230255f, 0.84485356524971f, 0.85772861000027f, 0.87008699110871f,
+    0.88192126434835f, 0.89322430119552f, 0.90398929312344f, 0.91420975570353f,
+    0.92387953251129f, 0.93299279883474f, 0.94154406518302f, 0.94952818059304f,
+    0.95694033573221f, 0.96377606579544f, 0.97003125319454f, 0.97570213003853f,
+    0.98078528040323f, 0.98527764238894f, 0.98917650996478f, 0.99247953459871f,
+    0.99518472667220f, 0.99729045667869f, 0.99879545620517f, 0.99969881869620f,
+    1.00000000000000f, 0.99969881869620f, 0.99879545620517f, 0.99729045667869f,
+    0.99518472667220f, 0.99247953459871f, 0.98917650996478f, 0.98527764238894f,
+    0.98078528040323f, 0.97570213003853f, 0.97003125319454f, 0.96377606579544f,
+    0.95694033573221f, 0.94952818059304f, 0.94154406518302f, 0.93299279883474f,
+    0.92387953251129f, 0.91420975570353f, 0.90398929312344f, 0.89322430119552f,
+    0.88192126434835f, 0.87008699110871f, 0.85772861000027f, 0.84485356524971f,
+    0.83146961230255f, 0.81758481315158f, 0.80320753148064f, 0.78834642762661f,
+    0.77301045336274f, 0.75720884650648f, 0.74095112535496f, 0.72424708295147f,
+    0.70710678118655f, 0.68954054473707f, 0.67155895484702f, 0.65317284295378f,
+    0.63439328416365f, 0.61523159058063f, 0.59569930449243f, 0.57580819141785f,
+    0.55557023301960f, 0.53499761988710f, 0.51410274419322f, 0.49289819222978f,
+    0.47139673682600f, 0.44961132965461f, 0.42755509343028f, 0.40524131400499f,
+    0.38268343236509f, 0.35989503653499f, 0.33688985339222f, 0.31368174039889f,
+    0.29028467725446f, 0.26671275747490f, 0.24298017990326f, 0.21910124015687f,
+    0.19509032201613f, 0.17096188876030f, 0.14673047445536f, 0.12241067519922f,
+    0.09801714032956f, 0.07356456359967f, 0.04906767432742f, 0.02454122852291f};
+
+}  // namespace
+
+SuppressionFilter::SuppressionFilter(int sample_rate_hz)
+    : sample_rate_hz_(sample_rate_hz),
+      e_output_old_(NumBandsForRate(sample_rate_hz_)) {
+  RTC_DCHECK(ValidFullBandRate(sample_rate_hz_));
+  e_input_old_.fill(0.f);
+  std::for_each(e_output_old_.begin(), e_output_old_.end(),
+                [](std::array<float, kFftLengthBy2>& a) { a.fill(0.f); });
+}
+
+SuppressionFilter::~SuppressionFilter() = default;
+
+void SuppressionFilter::ApplyGain(
+    const FftData& comfort_noise,
+    const FftData& comfort_noise_high_band,
+    const std::array<float, kFftLengthBy2Plus1>& suppression_gain,
+    std::vector<std::vector<float>>* e) {
+  RTC_DCHECK(e);
+  RTC_DCHECK_EQ(e->size(), NumBandsForRate(sample_rate_hz_));
+  FftData E;
+  std::array<float, kFftLength> e_extended;
+  constexpr float kIfftNormalization = 2.f / kFftLength;
+
+  // Analysis filterbank.
+  std::transform(e_input_old_.begin(), e_input_old_.end(),
+                 std::begin(kSqrtHanning), e_extended.begin(),
+                 std::multiplies<float>());
+  std::transform((*e)[0].begin(), (*e)[0].end(),
+                 std::begin(kSqrtHanning) + kFftLengthBy2,
+                 e_extended.begin() + kFftLengthBy2, std::multiplies<float>());
+  std::copy((*e)[0].begin(), (*e)[0].end(), e_input_old_.begin());
+  fft_.Fft(&e_extended, &E);
+
+  // Apply gain.
+  std::transform(suppression_gain.begin(), suppression_gain.end(), E.re.begin(),
+                 E.re.begin(), std::multiplies<float>());
+  std::transform(suppression_gain.begin(), suppression_gain.end(), E.im.begin(),
+                 E.im.begin(), std::multiplies<float>());
+
+  // Compute and add the comfort noise.
+  std::array<float, kFftLengthBy2Plus1> scaled_comfort_noise;
+  std::transform(suppression_gain.begin(), suppression_gain.end(),
+                 comfort_noise.re.begin(), scaled_comfort_noise.begin(),
+                 [](float a, float b) { return std::max(1.f - a, 0.f) * b; });
+  std::transform(scaled_comfort_noise.begin(), scaled_comfort_noise.end(),
+                 E.re.begin(), E.re.begin(), std::plus<float>());
+  std::transform(suppression_gain.begin(), suppression_gain.end(),
+                 comfort_noise.im.begin(), scaled_comfort_noise.begin(),
+                 [](float a, float b) { return std::max(1.f - a, 0.f) * b; });
+  std::transform(scaled_comfort_noise.begin(), scaled_comfort_noise.end(),
+                 E.im.begin(), E.im.begin(), std::plus<float>());
+
+  // Synthesis filterbank.
+  fft_.Ifft(E, &e_extended);
+  std::transform(e_output_old_[0].begin(), e_output_old_[0].end(),
+                 std::begin(kSqrtHanning) + kFftLengthBy2, (*e)[0].begin(),
+                 [&](float a, float b) { return kIfftNormalization * a * b; });
+  std::transform(e_extended.begin(), e_extended.begin() + kFftLengthBy2,
+                 std::begin(kSqrtHanning), e_extended.begin(),
+                 [&](float a, float b) { return kIfftNormalization * a * b; });
+  std::transform((*e)[0].begin(), (*e)[0].end(), e_extended.begin(),
+                 (*e)[0].begin(), std::plus<float>());
+  std::for_each((*e)[0].begin(), (*e)[0].end(), [](float& x_k) {
+    x_k = std::max(std::min(x_k, 32767.0f), -32768.0f);
+  });
+  std::copy(e_extended.begin() + kFftLengthBy2, e_extended.begin() + kFftLength,
+            std::begin(e_output_old_[0]));
+
+  if (e->size() > 1) {
+    // Form time-domain high-band noise.
+    std::array<float, kFftLength> time_domain_high_band_noise;
+    std::transform(comfort_noise_high_band.re.begin(),
+                   comfort_noise_high_band.re.end(), E.re.begin(),
+                   [&](float a) { return kIfftNormalization * a; });
+    std::transform(comfort_noise_high_band.im.begin(),
+                   comfort_noise_high_band.im.end(), E.im.begin(),
+                   [&](float a) { return kIfftNormalization * a; });
+    fft_.Ifft(E, &time_domain_high_band_noise);
+
+    // Scale and apply the noise to the signals.
+    // TODO(peah): Ensure that the high bands are properly delayed.
+    constexpr int kNumBandsAveragingUpperGain = kFftLengthBy2 / 4;
+    constexpr float kOneByNumBandsAveragingUpperGain =
+        1.f / kNumBandsAveragingUpperGain;
+    float high_bands_gain =
+        std::accumulate(suppression_gain.end() - kNumBandsAveragingUpperGain,
+                        suppression_gain.end(), 0.f) *
+        kOneByNumBandsAveragingUpperGain;
+
+    float high_bands_noise_scaling =
+        0.4f * std::max(1.f - high_bands_gain * high_bands_gain, 0.f);
+
+    std::transform(
+        (*e)[1].begin(), (*e)[1].end(), time_domain_high_band_noise.begin(),
+        (*e)[1].begin(), [&](float a, float b) {
+          return std::max(
+              std::min(b * high_bands_noise_scaling + high_bands_gain * a,
+                       32767.0f),
+              -32768.0f);
+        });
+
+    if (e->size() > 2) {
+      RTC_DCHECK_EQ(3, e->size());
+      std::for_each((*e)[2].begin(), (*e)[2].end(), [&](float& a) {
+        a = std::max(std::min(a * high_bands_gain, 32767.0f), -32768.0f);
+      });
+    }
+
+    std::array<float, kFftLengthBy2> tmp;
+    for (size_t k = 1; k < e->size(); ++k) {
+      std::copy((*e)[k].begin(), (*e)[k].end(), tmp.begin());
+      std::copy(e_output_old_[k].begin(), e_output_old_[k].end(),
+                (*e)[k].begin());
+      std::copy(tmp.begin(), tmp.end(), e_output_old_[k].begin());
+    }
+  }
+}
+
+}  // namespace webrtc