admin/webrtc_m130
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Fix the gain calculation in IntelligibilityEnhancer

Browse Source 
Review URL: https://codereview.webrtc.org/1718793002

Cr-Commit-Position: refs/heads/master@{#11755}
This commit is contained in:
aluebs 2016-02-24 17:25:42 -08:00 committed by Commit bot
parent 6140fcc11c
commit f99af6b885
3 changed files with 28 additions and 30 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

41
webrtc/modules/audio_processing/intelligibility/intelligibility_enhancer.cc
View File

@ -30,7 +30,7 @@ const int kChunkSizeMs = 10; // Size provided by APM.
const float kClipFreqKhz = 0.2f; const float kClipFreqKhz = 0.2f;
const float kKbdAlpha = 1.5f; const float kKbdAlpha = 1.5f;
const float kLambdaBot = -1.0f; // Extreme values in bisection const float kLambdaBot = -1.0f; // Extreme values in bisection
const float kLambdaTop = -10e-18f; // search for lamda. const float kLambdaTop = -1e-5f; // search for lamda.
const float kVoiceProbabilityThreshold = 0.02f; const float kVoiceProbabilityThreshold = 0.02f;
// Number of chunks after voice activity which is still considered speech. // Number of chunks after voice activity which is still considered speech.
const size_t kSpeechOffsetDelay = 80; const size_t kSpeechOffsetDelay = 80;
@ -164,15 +164,13 @@ void IntelligibilityEnhancer::ProcessClearBlock(
const float power_bot = const float power_bot =
DotProduct(gains_eq_.get(), filtered_clear_pow_.get(), bank_size_); DotProduct(gains_eq_.get(), filtered_clear_pow_.get(), bank_size_);
if (power_target >= power_bot && power_target <= power_top) { if (power_target >= power_bot && power_target <= power_top) {
SolveForLambda(power_target, power_bot, power_top); SolveForLambda(power_target);
UpdateErbGains(); UpdateErbGains();
} // Else experiencing power underflow, so do nothing. } // Else experiencing power underflow, so do nothing.
gain_applier_.Apply(in_block, out_block); gain_applier_.Apply(in_block, out_block);
} }
void IntelligibilityEnhancer::SolveForLambda(float power_target, void IntelligibilityEnhancer::SolveForLambda(float power_target) {
float power_bot,
float power_top) {
const float kConvergeThresh = 0.001f; // TODO(ekmeyerson): Find best values const float kConvergeThresh = 0.001f; // TODO(ekmeyerson): Find best values
const int kMaxIters = 100; // for these, based on experiments. const int kMaxIters = 100; // for these, based on experiments.
@ -183,7 +181,7 @@ void IntelligibilityEnhancer::SolveForLambda(float power_target,
float power_ratio = 2.f; // Ratio of achieved power to target power. float power_ratio = 2.f; // Ratio of achieved power to target power.
int iters = 0; int iters = 0;
while (std::fabs(power_ratio - 1.f) > kConvergeThresh && iters <= kMaxIters) { while (std::fabs(power_ratio - 1.f) > kConvergeThresh && iters <= kMaxIters) {
const float lambda = lambda_bot + (lambda_top - lambda_bot) / 2.f; const float lambda = (lambda_bot + lambda_top) / 2.f;
SolveForGainsGivenLambda(lambda, start_freq_, gains_eq_.get()); SolveForGainsGivenLambda(lambda, start_freq_, gains_eq_.get());
const float power = const float power =
DotProduct(gains_eq_.get(), filtered_clear_pow_.get(), bank_size_); DotProduct(gains_eq_.get(), filtered_clear_pow_.get(), bank_size_);
@ -286,7 +284,8 @@ std::vector<std::vector<float>> IntelligibilityEnhancer::CreateErbBank(
void IntelligibilityEnhancer::SolveForGainsGivenLambda(float lambda, void IntelligibilityEnhancer::SolveForGainsGivenLambda(float lambda,
size_t start_freq, size_t start_freq,
float* sols) { float* sols) {
bool quadratic = (kRho < 1.f); const float kMinPower = 1e-5f;
const float* pow_x0 = filtered_clear_pow_.get(); const float* pow_x0 = filtered_clear_pow_.get();
const float* pow_n0 = filtered_noise_pow_.get(); const float* pow_n0 = filtered_noise_pow_.get();
@ -295,20 +294,24 @@ void IntelligibilityEnhancer::SolveForGainsGivenLambda(float lambda,
} }
// Analytic solution for optimal gains. See paper for derivation. // Analytic solution for optimal gains. See paper for derivation.
for (size_t n = start_freq - 1; n < bank_size_; ++n) { for (size_t n = start_freq; n < bank_size_; ++n) {
float alpha0, beta0, gamma0; if (pow_x0[n] < kMinPower || pow_n0[n] < kMinPower) {
gamma0 = 0.5f * kRho * pow_x0[n] * pow_n0[n] + sols[n] = 1.f;
lambda * pow_x0[n] * pow_n0[n] * pow_n0[n];
beta0 = lambda * pow_x0[n] * (2 - kRho) * pow_x0[n] * pow_n0[n];
if (quadratic) {
alpha0 = lambda * pow_x0[n] * (1 - kRho) * pow_x0[n] * pow_x0[n];
sols[n] =
(-beta0 - sqrtf(beta0 * beta0 - 4 * alpha0 * gamma0)) /
(2 * alpha0 + std::numeric_limits<float>::epsilon());
} else { } else {
sols[n] = -gamma0 / beta0; const float gamma0 = 0.5f * kRho * pow_x0[n] * pow_n0[n] +
lambda * pow_x0[n] * pow_n0[n] * pow_n0[n];
const float beta0 =
lambda * pow_x0[n] * (2.f - kRho) * pow_x0[n] * pow_n0[n];
const float alpha0 =
lambda * pow_x0[n] * (1.f - kRho) * pow_x0[n] * pow_x0[n];
RTC_DCHECK_LT(alpha0, 0.f);
// The quadratic equation should always have real roots, but to guard
// against numerical errors we limit it to a minimum of zero.
sols[n] = std::max(
0.f, (-beta0 - std::sqrt(std::max(
0.f, beta0 * beta0 - 4.f * alpha0 * gamma0))) /
(2.f * alpha0));
} }
sols[n] = fmax(0, sols[n]);
} }
} }

2
webrtc/modules/audio_processing/intelligibility/intelligibility_enhancer.h
View File

@ -67,7 +67,7 @@ class IntelligibilityEnhancer {
std::complex<float>* out_block); std::complex<float>* out_block);
// Bisection search for optimal |lambda|. // Bisection search for optimal |lambda|.
void SolveForLambda(float power_target, float power_bot, float power_top); void SolveForLambda(float power_target);
// Transforms freq gains to ERB gains. // Transforms freq gains to ERB gains.
void UpdateErbGains(); void UpdateErbGains();

15
webrtc/modules/audio_processing/intelligibility/intelligibility_enhancer_unittest.cc
View File

@ -186,16 +186,11 @@ static_assert(arraysize(kTestCenterFreqs) == arraysize(kTestFilterBank),
// Target output for gain solving test. Generated with matlab. // Target output for gain solving test. Generated with matlab.
const size_t kTestStartFreq = 12; // Lowest integral frequency for ERBs. const size_t kTestStartFreq = 12; // Lowest integral frequency for ERBs.
const float kTestZeroVar[] = { const float kTestZeroVar = 1.f;
1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 0.f, 0.f, 0.f,
0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f,
0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0};
static_assert(arraysize(kTestCenterFreqs) == arraysize(kTestZeroVar),
"Power test data badly initialized.");
const float kTestNonZeroVarLambdaTop[] = { const float kTestNonZeroVarLambdaTop[] = {
1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 0.f, 0.f, 0.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 0.f, 0.f,
0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f,
0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0}; 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f};
static_assert(arraysize(kTestCenterFreqs) == static_assert(arraysize(kTestCenterFreqs) ==
arraysize(kTestNonZeroVarLambdaTop), arraysize(kTestNonZeroVarLambdaTop),
"Power test data badly initialized."); "Power test data badly initialized.");
@ -280,7 +275,7 @@ TEST_F(IntelligibilityEnhancerTest, TestSolveForGains) {
} }
enh_->SolveForGainsGivenLambda(lambda, enh_->start_freq_, &sols[0]); enh_->SolveForGainsGivenLambda(lambda, enh_->start_freq_, &sols[0]);
for (size_t i = 0; i < enh_->bank_size_; i++) { for (size_t i = 0; i < enh_->bank_size_; i++) {
EXPECT_NEAR(kTestZeroVar[i], sols[i], kMaxTestError); EXPECT_NEAR(kTestZeroVar, sols[i], kMaxTestError);
} }
for (size_t i = 0; i < enh_->bank_size_; i++) { for (size_t i = 0; i < enh_->bank_size_; i++) {
enh_->filtered_clear_pow_[i] = static_cast<float>(i + 1); enh_->filtered_clear_pow_[i] = static_cast<float>(i + 1);
@ -293,7 +288,7 @@ TEST_F(IntelligibilityEnhancerTest, TestSolveForGains) {
lambda = -1.f; lambda = -1.f;
enh_->SolveForGainsGivenLambda(lambda, enh_->start_freq_, &sols[0]); enh_->SolveForGainsGivenLambda(lambda, enh_->start_freq_, &sols[0]);
for (size_t i = 0; i < enh_->bank_size_; i++) { for (size_t i = 0; i < enh_->bank_size_; i++) {
EXPECT_NEAR(kTestZeroVar[i], sols[i], kMaxTestError); EXPECT_NEAR(kTestNonZeroVarLambdaTop[i], sols[i], kMaxTestError);
} }
} }