Add new variance update option and unittests for intelligibility

webrtc/modules/audio_processing/intelligibility/test/enhancer_unittest.cc

Index: webrtc/modules/audio_processing/intelligibility/test/enhancer_unittest.cc
diff --git a/webrtc/modules/audio_processing/intelligibility/test/enhancer_unittest.cc b/webrtc/modules/audio_processing/intelligibility/test/enhancer_unittest.cc
new file mode 100644
index 0000000000000000000000000000000000000000..169234c643a64af5798631d62fa6592be60c69cd
--- /dev/null
+++ b/webrtc/modules/audio_processing/intelligibility/test/enhancer_unittest.cc
@@ -0,0 +1,201 @@
+/*
+ *  Copyright (c) 2014 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.
+ */
+
+//
+//  Unit tests for intelligibility enhancer.
+//
+
+#include <cmath>
+#include <algorithm>
+
+#include "testing/gtest/include/gtest/gtest.h"
+#include "webrtc/common_audio/signal_processing/include/signal_processing_library.h"
+#include "webrtc/modules/audio_processing/intelligibility/intelligibility_enhancer.h"
+
+using std::vector;
+using webrtc::intelligibility::VarianceArray;
+
+namespace webrtc {
+
+// Generated with matlab code: normrnd(0,1000,64,1).
+const double kGaussianSamples[64] = {1689.1, 1437, -2251.1, 356.49, -850.24,
+    -299.55, -634.25, 1624.5, 1241.1, 555.28, 703.42, 458.16, 683.98, 251.29,
+    -178.5, 507.73, -309.9, -394.37, -269.74, -88.13, 8.0293, 2531.8, -1223.2,
+    -1071.8, 246.06, -50.611, -730.15, 326.99, 752.99, -1153.7, -407.87,
+    -1287.9, 83.578, 163.8, 682.57, -1086.4, 297.49, -143.31, 1392, 306.75,
+    -537.18, -228.93, -536.22, 1439, -511.1, -1606.8, -201.24, 1143.5, 663.29,
+     164.08, 1785.4, -587.71, 259.04, -871.83, -787.92, -344.34, 647.62,
+     2054.1, 798.94, -1071.1, -205.16, -554.44, -292.94, 1180.2};
+
+// Target output for ERB create test. Generated with matlab.
+const double kTestNumCenterFreqs = 22;
+const double kTestCenterFreqs[22] = {13.169, 26.965, 41.423, 56.577, 72.461,
+    89.113, 106.57, 124.88, 144.08, 164.21, 185.34, 207.5, 230.75, 255.16,
+    280.77, 307.66, 335.9, 365.56, 396.71, 429.44, 463.84, 500};
+const double kTestNumFreqs = 2;
+const double kTestFilterBank[22][2] = { {0.055556, 0}, {0.055556, 0},
+                                        {0.055556, 0}, {0.055556, 0},
+                                        {0.055556, 0}, {0.055556, 0},
+                                        {0.055556, 0}, {0.055556, 0},
+                                        {0.055556, 0}, {0.055556, 0},
+                                        {0.055556, 0}, {0.055556, 0},
+                                        {0.055556, 0}, {0.055556, 0},
+                                        {0.055556, 0}, {0.055556, 0},
+                                        {0.055556, 0}, {0.055556, 0.2},
+                                        {0, 0.2}, {0, 0.2},
+                                        {0, 0.2}, {0, 0.2} };
+// Target output for gain solving test. Generated with matlab.
+const double kTestZeroVar[22] = {1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+                                 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
+const double kTestNonZeroVarLambdaTop[22] = {1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+                                             0, 0, 0.0351, 0.0636, 0.0863,
+                                             0.1037, 0.1162, 0.1236, 0.1251,
+                                             0.1189, 0.0993};
+const float kMaxTestError = 0.005;
+
+// Enhancer initialization parameters.
+const int kSamples = 2000;
+const int kErbResolution = 2;
+const int kSampleRate = 1000;
+const int kFragmentSize = kSampleRate / 100;
+const int kNumChannels = 1;
+const float kDecayRate = 0.9f;
+const int kWindowSize = 800;
+const int kAnalyzeRate = 800;
+const int kVarianceRate = 2;
+const float kGainLimit = 0.1f;
+
+void GenerateConstantData(vector<float>& data, float constant) {
+  for (size_t i = 0; i < data.size(); i++) {
+    data[i] = constant;
+  }
+}
+
+void GenerateGaussianData(vector<float>& data) {
+  static int count = 0;
+  for (size_t i = 0; i < data.size(); i++) {
+    data[i] = kGaussianSamples[count%64];
+    count++;
+  }
+}
+
+
+class EnhancerTest : public ::testing::Test {
+ protected:
+  IntelligibilityEnhancer enh_;
+  vector<float> clear_data_;
+  vector<float> noise_data_;
+  EnhancerTest() :
+      enh_(kErbResolution,
+           kSampleRate,
+           kNumChannels,
+           VarianceArray::kStepInfinite,
+           kDecayRate,
+           kWindowSize,
+           kAnalyzeRate,
+           kVarianceRate,
+           kGainLimit),
+      clear_data_(kSamples),
+      noise_data_(kSamples) {}
+
+  void RunEnhancer(VarianceArray::StepType step_type) {

[turaj, 2015/06/26 00:32:58]

Could this function use |enh_| instead?

[ekm, 2015/06/26 19:07:09]

That would be better, but I was having trouble modifying |step_type| after the
first construction without using placement new, or reaching deep into
VarianceArray. Is there a nice way to do this? Could do something like give
enhancer a SetVarianceType method.

[turaj, 2015/06/29 17:33:35]

Sorry, I didn't notice that |step_type| is an input. I agree that it does not
worth to modify the code just for this test. One other way that the constructor
takes an input, but I guess it is fine as it is.

+    IntelligibilityEnhancer enh(kErbResolution,
+                                kSampleRate,
+                                kNumChannels,
+                                step_type,
+                                kDecayRate,
+                                kWindowSize,
+                                kAnalyzeRate,
+                                kVarianceRate,
+                                kGainLimit);
+    float* clear_cursor = &clear_data_[0];
+    float* noise_cursor = &noise_data_[0];
+    for (int i = 0; i < kSamples; i+= kFragmentSize) {
+      enh.ProcessCaptureAudio(&noise_cursor);
+      enh.ProcessRenderAudio(&clear_cursor);
+      clear_cursor += kFragmentSize;
+      noise_cursor += kFragmentSize;
+    }
+  }
+};
+
+// For each class of generated data, tests plumbing for
+// each variance update method.
+TEST_F(EnhancerTest, TestPlumbing) {

[turaj, 2015/06/26 00:32:58]

what is it that ids tested here? How could it fail? If you like to test pluming,
you have to inject a mock of variance updater and expect certain calls to
happen. 

[ekm, 2015/06/26 19:07:09]

Before we switched to doing nothing in case of underflow, some of the variance
array types were raising the error, and this helped to see when that happened.
Maybe now this should test that ProcessRenderAudio modifies the render stream
when we expect it to? The mock may be useful for checking when Step is called
when we're using VAD, but I don't see what it'd be useful for now.

[turaj, 2015/06/29 17:33:36]

Agreed, this is a better test. 

+  vector<VarianceArray::StepType> step_types = {
+      VarianceArray::kStepInfinite, VarianceArray::kStepDecaying,
+      VarianceArray::kStepWindowed, VarianceArray::kStepBlocked,
+      VarianceArray::kStepBlockBasedMovingAverage};
+  for (vector<VarianceArray::StepType>::iterator step_type =
+      step_types.begin(); step_type != step_types.end(); ++step_type) {
+    GenerateConstantData(clear_data_, 0.0f);
+    GenerateConstantData(noise_data_, 0.0f);
+    RunEnhancer(*step_type);
+    GenerateConstantData(clear_data_, 500.0f);
+    RunEnhancer(*step_type);
+    GenerateConstantData(noise_data_, 500.0f);
+    RunEnhancer(*step_type);
+    GenerateGaussianData(clear_data_);
+    RunEnhancer(*step_type);
+    GenerateGaussianData(noise_data_);
+    RunEnhancer(*step_type);
+    GenerateConstantData(clear_data_, 0);
+    RunEnhancer(*step_type);
+  }
+}
+
+// Tests ERB bank creation, comparing against matlab output.
+TEST_F(EnhancerTest, TestErbCreation) {
+  ASSERT_EQ(enh_.bank_size_, kTestNumCenterFreqs);
+  for (int i = 0; i < enh_.bank_size_; ++i) {
+    float error = std::fabs(enh_.center_freqs_[i] - kTestCenterFreqs[i]);
+    EXPECT_LT(error, kMaxTestError);
+    ASSERT_EQ(enh_.freqs_, kTestNumFreqs);
+    for (int j = 0; j < enh_.freqs_; ++j) {
+      float error = std::fabs(enh_.filter_bank_[i][j] - kTestFilterBank[i][j]);
+      EXPECT_LT(error, kMaxTestError);
+    }
+  }
+}
+
+// Tests analytic solution for optimal gains, comparing
+// against matlab output.
+TEST_F(EnhancerTest, TestSolveForGains) {
+  ASSERT_EQ(enh_.start_freq_, 12);
+  vector<float> sols(enh_.bank_size_);
+  float lambda = -0.001;

[turaj, 2015/06/26 00:32:58]

I guess you need -0.001f otherwise Visual Studio complains.

[ekm, 2015/06/26 19:07:09]

Done.

+  for (int i = 0; i < enh_.bank_size_; i++) {
+    enh_.filtered_clear_var_[i] = 0.0;
+    enh_.filtered_noise_var_[i] = 0.0;
+    enh_.rho_[i] = 0.02;
+  }
+  enh_.SolveForGainsGivenLambda(lambda, enh_.start_freq_, &sols[0]);
+  for (int i = 0; i < enh_.bank_size_; i++) {
+    float error = std::fabs(sols[i] - kTestZeroVar[i]);
+    EXPECT_LT(error, kMaxTestError);

[turaj, 2015/06/26 00:32:58]

I guess you can use EXPECT_NEAR(v1, v2, tolerance) here and later and on line
163.

[ekm, 2015/06/26 19:07:10]

Done.

+  }
+  for (int i = 0; i < enh_.bank_size_; i++) {
+    enh_.filtered_clear_var_[i] = static_cast<float>(i+1);

[turaj, 2015/06/26 00:32:58]

'i + 1'

[ekm, 2015/06/26 19:07:10]

Done.

+    enh_.filtered_noise_var_[i] = static_cast<float>(enh_.bank_size_ - i);
+  }
+  enh_.SolveForGainsGivenLambda(lambda, enh_.start_freq_, &sols[0]);
+  for (int i = 0; i < enh_.bank_size_; i++) {
+    float error = std::fabs(sols[i] - kTestNonZeroVarLambdaTop[i]);
+    EXPECT_LT(error, kMaxTestError);
+  }
+  lambda = -1.0;
+  enh_.SolveForGainsGivenLambda(lambda, enh_.start_freq_, &sols[0]);
+  for (int i = 0; i < enh_.bank_size_; i++) {
+    float error = std::fabs(sols[i] - kTestZeroVar[i]);
+    EXPECT_LT(error, kMaxTestError);
+  }
+}
+
+}  // namespace webrtc