Fix and simplify the power estimation in the IntelligibilityEnhancer
R=henrik.lundin@webrtc.org, turaj@webrtc.org Review URL: https://codereview.webrtc.org/1685703004 . Cr-Commit-Position: refs/heads/master@{#11663}
This commit is contained in:
Alejandro Luebs
parent
ee18220ddd
commit
32348192cc
@ -8,13 +8,6 @@
|
||||
* be found in the AUTHORS file in the root of the source tree.
|
||||
*/
|
||||
|
||||
//
|
||||
// Implements core class for intelligibility enhancer.
|
||||
//
|
||||
// Details of the model and algorithm can be found in the original paper:
|
||||
// http://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=6882788
|
||||
//
|
||||
|
||||
#include "webrtc/modules/audio_processing/intelligibility/intelligibility_enhancer.h"
|
||||
|
||||
#include <math.h>
|
||||
@ -32,7 +25,7 @@ namespace webrtc {
|
||||
namespace {
|
||||
|
||||
const size_t kErbResolution = 2;
|
||||
const int kWindowSizeMs = 2;
|
||||
const int kWindowSizeMs = 16;
|
||||
const int kChunkSizeMs = 10; // Size provided by APM.
|
||||
const float kClipFreq = 200.0f;
|
||||
const float kConfigRho = 0.02f; // Default production and interpretation SNR.
|
||||
@ -49,35 +42,30 @@ float DotProduct(const float* a, const float* b, size_t length) {
|
||||
return ret;
|
||||
}
|
||||
|
||||
// Computes the power across ERB filters from the power spectral density |var|.
|
||||
// Computes the power across ERB bands from the power spectral density |pow|.
|
||||
// Stores it in |result|.
|
||||
void FilterVariance(const float* var,
|
||||
const std::vector<std::vector<float>>& filter_bank,
|
||||
float* result) {
|
||||
void MapToErbBands(const float* pow,
|
||||
const std::vector<std::vector<float>>& filter_bank,
|
||||
float* result) {
|
||||
for (size_t i = 0; i < filter_bank.size(); ++i) {
|
||||
RTC_DCHECK_GT(filter_bank[i].size(), 0u);
|
||||
result[i] = DotProduct(&filter_bank[i][0], var, filter_bank[i].size());
|
||||
result[i] = DotProduct(&filter_bank[i][0], pow, filter_bank[i].size());
|
||||
}
|
||||
}
|
||||
|
||||
} // namespace
|
||||
|
||||
using std::complex;
|
||||
using std::max;
|
||||
using std::min;
|
||||
using VarianceType = intelligibility::VarianceArray::StepType;
|
||||
|
||||
IntelligibilityEnhancer::TransformCallback::TransformCallback(
|
||||
IntelligibilityEnhancer* parent)
|
||||
: parent_(parent) {
|
||||
}
|
||||
|
||||
void IntelligibilityEnhancer::TransformCallback::ProcessAudioBlock(
|
||||
const complex<float>* const* in_block,
|
||||
const std::complex<float>* const* in_block,
|
||||
size_t in_channels,
|
||||
size_t frames,
|
||||
size_t /* out_channels */,
|
||||
complex<float>* const* out_block) {
|
||||
std::complex<float>* const* out_block) {
|
||||
RTC_DCHECK_EQ(parent_->freqs_, frames);
|
||||
for (size_t i = 0; i < in_channels; ++i) {
|
||||
parent_->ProcessClearBlock(in_block[i], out_block[i]);
|
||||
@ -101,13 +89,10 @@ IntelligibilityEnhancer::IntelligibilityEnhancer(const Config& config)
|
||||
num_render_channels_(config.num_render_channels),
|
||||
analysis_rate_(config.analysis_rate),
|
||||
active_(true),
|
||||
clear_variance_(freqs_,
|
||||
config.var_type,
|
||||
config.var_window_size,
|
||||
config.var_decay_rate),
|
||||
clear_power_(freqs_, config.decay_rate),
|
||||
noise_power_(freqs_, 0.f),
|
||||
filtered_clear_var_(new float[bank_size_]),
|
||||
filtered_noise_var_(new float[bank_size_]),
|
||||
filtered_clear_pow_(new float[bank_size_]),
|
||||
filtered_noise_pow_(new float[bank_size_]),
|
||||
center_freqs_(new float[bank_size_]),
|
||||
render_filter_bank_(CreateErbBank(freqs_)),
|
||||
rho_(new float[bank_size_]),
|
||||
@ -120,12 +105,12 @@ IntelligibilityEnhancer::IntelligibilityEnhancer(const Config& config)
|
||||
analysis_step_(0) {
|
||||
RTC_DCHECK_LE(config.rho, 1.0f);
|
||||
|
||||
memset(filtered_clear_var_.get(),
|
||||
memset(filtered_clear_pow_.get(),
|
||||
0,
|
||||
bank_size_ * sizeof(filtered_clear_var_[0]));
|
||||
memset(filtered_noise_var_.get(),
|
||||
bank_size_ * sizeof(filtered_clear_pow_[0]));
|
||||
memset(filtered_noise_pow_.get(),
|
||||
0,
|
||||
bank_size_ * sizeof(filtered_noise_var_[0]));
|
||||
bank_size_ * sizeof(filtered_noise_pow_[0]));
|
||||
|
||||
// Assumes all rho equal.
|
||||
for (size_t i = 0; i < bank_size_; ++i) {
|
||||
@ -176,8 +161,9 @@ void IntelligibilityEnhancer::ProcessRenderAudio(float* const* audio,
|
||||
}
|
||||
}
|
||||
|
||||
void IntelligibilityEnhancer::ProcessClearBlock(const complex<float>* in_block,
|
||||
complex<float>* out_block) {
|
||||
void IntelligibilityEnhancer::ProcessClearBlock(
|
||||
const std::complex<float>* in_block,
|
||||
std::complex<float>* out_block) {
|
||||
if (block_count_ < 2) {
|
||||
memset(out_block, 0, freqs_ * sizeof(*out_block));
|
||||
++block_count_;
|
||||
@ -186,11 +172,9 @@ void IntelligibilityEnhancer::ProcessClearBlock(const complex<float>* in_block,
|
||||
|
||||
// TODO(ekm): Use VAD to |Step| and |AnalyzeClearBlock| only if necessary.
|
||||
if (true) {
|
||||
clear_variance_.Step(in_block, false);
|
||||
clear_power_.Step(in_block);
|
||||
if (block_count_ % analysis_rate_ == analysis_rate_ - 1) {
|
||||
const float power_target = std::accumulate(
|
||||
clear_variance_.variance(), clear_variance_.variance() + freqs_, 0.f);
|
||||
AnalyzeClearBlock(power_target);
|
||||
AnalyzeClearBlock();
|
||||
++analysis_step_;
|
||||
}
|
||||
++block_count_;
|
||||
@ -201,23 +185,26 @@ void IntelligibilityEnhancer::ProcessClearBlock(const complex<float>* in_block,
|
||||
}
|
||||
}
|
||||
|
||||
void IntelligibilityEnhancer::AnalyzeClearBlock(float power_target) {
|
||||
FilterVariance(clear_variance_.variance(),
|
||||
render_filter_bank_,
|
||||
filtered_clear_var_.get());
|
||||
FilterVariance(&noise_power_[0],
|
||||
capture_filter_bank_,
|
||||
filtered_noise_var_.get());
|
||||
void IntelligibilityEnhancer::AnalyzeClearBlock() {
|
||||
const float* clear_power = clear_power_.Power();
|
||||
MapToErbBands(clear_power,
|
||||
render_filter_bank_,
|
||||
filtered_clear_pow_.get());
|
||||
MapToErbBands(&noise_power_[0],
|
||||
capture_filter_bank_,
|
||||
filtered_noise_pow_.get());
|
||||
SolveForGainsGivenLambda(kLambdaTop, start_freq_, gains_eq_.get());
|
||||
const float power_target = std::accumulate(
|
||||
clear_power, clear_power + freqs_, 0.f);
|
||||
const float power_top =
|
||||
DotProduct(gains_eq_.get(), filtered_clear_var_.get(), bank_size_);
|
||||
DotProduct(gains_eq_.get(), filtered_clear_pow_.get(), bank_size_);
|
||||
SolveForGainsGivenLambda(kLambdaBot, start_freq_, gains_eq_.get());
|
||||
const float power_bot =
|
||||
DotProduct(gains_eq_.get(), filtered_clear_var_.get(), bank_size_);
|
||||
DotProduct(gains_eq_.get(), filtered_clear_pow_.get(), bank_size_);
|
||||
if (power_target >= power_bot && power_target <= power_top) {
|
||||
SolveForLambda(power_target, power_bot, power_top);
|
||||
UpdateErbGains();
|
||||
} // Else experiencing variance underflow, so do nothing.
|
||||
} // Else experiencing power underflow, so do nothing.
|
||||
}
|
||||
|
||||
void IntelligibilityEnhancer::SolveForLambda(float power_target,
|
||||
@ -237,7 +224,7 @@ void IntelligibilityEnhancer::SolveForLambda(float power_target,
|
||||
const float lambda = lambda_bot + (lambda_top - lambda_bot) / 2.0f;
|
||||
SolveForGainsGivenLambda(lambda, start_freq_, gains_eq_.get());
|
||||
const float power =
|
||||
DotProduct(gains_eq_.get(), filtered_clear_var_.get(), bank_size_);
|
||||
DotProduct(gains_eq_.get(), filtered_clear_pow_.get(), bank_size_);
|
||||
if (power < power_target) {
|
||||
lambda_bot = lambda;
|
||||
} else {
|
||||
@ -290,22 +277,22 @@ std::vector<std::vector<float>> IntelligibilityEnhancer::CreateErbBank(
|
||||
size_t lll, ll, rr, rrr;
|
||||
static const size_t kOne = 1; // Avoids repeated static_cast<>s below.
|
||||
lll = static_cast<size_t>(round(
|
||||
center_freqs_[max(kOne, i - lf) - 1] * num_freqs /
|
||||
center_freqs_[std::max(kOne, i - lf) - 1] * num_freqs /
|
||||
(0.5f * sample_rate_hz_)));
|
||||
ll = static_cast<size_t>(round(
|
||||
center_freqs_[max(kOne, i) - 1] * num_freqs /
|
||||
center_freqs_[std::max(kOne, i) - 1] * num_freqs /
|
||||
(0.5f * sample_rate_hz_)));
|
||||
lll = min(num_freqs, max(lll, kOne)) - 1;
|
||||
ll = min(num_freqs, max(ll, kOne)) - 1;
|
||||
lll = std::min(num_freqs, std::max(lll, kOne)) - 1;
|
||||
ll = std::min(num_freqs, std::max(ll, kOne)) - 1;
|
||||
|
||||
rrr = static_cast<size_t>(round(
|
||||
center_freqs_[min(bank_size_, i + rf) - 1] * num_freqs /
|
||||
center_freqs_[std::min(bank_size_, i + rf) - 1] * num_freqs /
|
||||
(0.5f * sample_rate_hz_)));
|
||||
rr = static_cast<size_t>(round(
|
||||
center_freqs_[min(bank_size_, i + 1) - 1] * num_freqs /
|
||||
center_freqs_[std::min(bank_size_, i + 1) - 1] * num_freqs /
|
||||
(0.5f * sample_rate_hz_)));
|
||||
rrr = min(num_freqs, max(rrr, kOne)) - 1;
|
||||
rr = min(num_freqs, max(rr, kOne)) - 1;
|
||||
rrr = std::min(num_freqs, std::max(rrr, kOne)) - 1;
|
||||
rr = std::min(num_freqs, std::max(rr, kOne)) - 1;
|
||||
|
||||
float step, element;
|
||||
|
||||
@ -343,8 +330,8 @@ void IntelligibilityEnhancer::SolveForGainsGivenLambda(float lambda,
|
||||
size_t start_freq,
|
||||
float* sols) {
|
||||
bool quadratic = (kConfigRho < 1.0f);
|
||||
const float* var_x0 = filtered_clear_var_.get();
|
||||
const float* var_n0 = filtered_noise_var_.get();
|
||||
const float* pow_x0 = filtered_clear_pow_.get();
|
||||
const float* pow_n0 = filtered_noise_pow_.get();
|
||||
|
||||
for (size_t n = 0; n < start_freq; ++n) {
|
||||
sols[n] = 1.0f;
|
||||
@ -353,11 +340,11 @@ void IntelligibilityEnhancer::SolveForGainsGivenLambda(float lambda,
|
||||
// Analytic solution for optimal gains. See paper for derivation.
|
||||
for (size_t n = start_freq - 1; n < bank_size_; ++n) {
|
||||
float alpha0, beta0, gamma0;
|
||||
gamma0 = 0.5f * rho_[n] * var_x0[n] * var_n0[n] +
|
||||
lambda * var_x0[n] * var_n0[n] * var_n0[n];
|
||||
beta0 = lambda * var_x0[n] * (2 - rho_[n]) * var_x0[n] * var_n0[n];
|
||||
gamma0 = 0.5f * rho_[n] * pow_x0[n] * pow_n0[n] +
|
||||
lambda * pow_x0[n] * pow_n0[n] * pow_n0[n];
|
||||
beta0 = lambda * pow_x0[n] * (2 - rho_[n]) * pow_x0[n] * pow_n0[n];
|
||||
if (quadratic) {
|
||||
alpha0 = lambda * var_x0[n] * (1 - rho_[n]) * var_x0[n] * var_x0[n];
|
||||
alpha0 = lambda * pow_x0[n] * (1 - rho_[n]) * pow_x0[n] * pow_x0[n];
|
||||
sols[n] =
|
||||
(-beta0 - sqrtf(beta0 * beta0 - 4 * alpha0 * gamma0)) /
|
||||
(2 * alpha0 + std::numeric_limits<float>::epsilon());
|
||||
|
||||
@ -8,10 +8,6 @@
|
||||
* be found in the AUTHORS file in the root of the source tree.
|
||||
*/
|
||||
|
||||
//
|
||||
// Specifies core class for intelligbility enhancement.
|
||||
//
|
||||
|
||||
#ifndef WEBRTC_MODULES_AUDIO_PROCESSING_INTELLIGIBILITY_INTELLIGIBILITY_ENHANCER_H_
|
||||
#define WEBRTC_MODULES_AUDIO_PROCESSING_INTELLIGIBILITY_INTELLIGIBILITY_ENHANCER_H_
|
||||
|
||||
@ -28,30 +24,25 @@ namespace webrtc {
|
||||
// Speech intelligibility enhancement module. Reads render and capture
|
||||
// audio streams and modifies the render stream with a set of gains per
|
||||
// frequency bin to enhance speech against the noise background.
|
||||
// Note: assumes speech and noise streams are already separated.
|
||||
// Details of the model and algorithm can be found in the original paper:
|
||||
// http://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=6882788
|
||||
class IntelligibilityEnhancer {
|
||||
public:
|
||||
struct Config {
|
||||
// |var_*| are parameters for the VarianceArray constructor for the
|
||||
// clear speech stream.
|
||||
// TODO(bercic): the |var_*|, |*_rate| and |gain_limit| parameters should
|
||||
// probably go away once fine tuning is done.
|
||||
// TODO(bercic): the |decay_rate|, |analysis_rate| and |gain_limit|
|
||||
// parameters should probably go away once fine tuning is done.
|
||||
Config()
|
||||
: sample_rate_hz(16000),
|
||||
num_capture_channels(1),
|
||||
num_render_channels(1),
|
||||
var_type(intelligibility::VarianceArray::kStepDecaying),
|
||||
var_decay_rate(0.9f),
|
||||
var_window_size(10),
|
||||
analysis_rate(800),
|
||||
decay_rate(0.9f),
|
||||
analysis_rate(60),
|
||||
gain_change_limit(0.1f),
|
||||
rho(0.02f) {}
|
||||
int sample_rate_hz;
|
||||
size_t num_capture_channels;
|
||||
size_t num_render_channels;
|
||||
intelligibility::VarianceArray::StepType var_type;
|
||||
float var_decay_rate;
|
||||
size_t var_window_size;
|
||||
float decay_rate;
|
||||
int analysis_rate;
|
||||
float gain_change_limit;
|
||||
float rho;
|
||||
@ -90,13 +81,13 @@ class IntelligibilityEnhancer {
|
||||
FRIEND_TEST_ALL_PREFIXES(IntelligibilityEnhancerTest, TestErbCreation);
|
||||
FRIEND_TEST_ALL_PREFIXES(IntelligibilityEnhancerTest, TestSolveForGains);
|
||||
|
||||
// Updates variance computation and analysis with |in_block_|,
|
||||
// Updates power computation and analysis with |in_block_|,
|
||||
// and writes modified speech to |out_block|.
|
||||
void ProcessClearBlock(const std::complex<float>* in_block,
|
||||
std::complex<float>* out_block);
|
||||
|
||||
// Computes and sets modified gains.
|
||||
void AnalyzeClearBlock(float power_target);
|
||||
void AnalyzeClearBlock();
|
||||
|
||||
// Bisection search for optimal |lambda|.
|
||||
void SolveForLambda(float power_target, float power_bot, float power_top);
|
||||
@ -127,10 +118,10 @@ class IntelligibilityEnhancer {
|
||||
const bool active_; // Whether render gains are being updated.
|
||||
// TODO(ekm): Add logic for updating |active_|.
|
||||
|
||||
intelligibility::VarianceArray clear_variance_;
|
||||
intelligibility::PowerEstimator clear_power_;
|
||||
std::vector<float> noise_power_;
|
||||
rtc::scoped_ptr<float[]> filtered_clear_var_;
|
||||
rtc::scoped_ptr<float[]> filtered_noise_var_;
|
||||
rtc::scoped_ptr<float[]> filtered_clear_pow_;
|
||||
rtc::scoped_ptr<float[]> filtered_noise_pow_;
|
||||
rtc::scoped_ptr<float[]> center_freqs_;
|
||||
std::vector<std::vector<float>> capture_filter_bank_;
|
||||
std::vector<std::vector<float>> render_filter_bank_;
|
||||
|
||||
@ -8,10 +8,6 @@
|
||||
* be found in the AUTHORS file in the root of the source tree.
|
||||
*/
|
||||
|
||||
//
|
||||
// Unit tests for intelligibility enhancer.
|
||||
//
|
||||
|
||||
#include <math.h>
|
||||
#include <stdlib.h>
|
||||
#include <algorithm>
|
||||
@ -32,28 +28,29 @@ const float kTestCenterFreqs[] = {
|
||||
13.169f, 26.965f, 41.423f, 56.577f, 72.461f, 89.113f, 106.57f, 124.88f,
|
||||
144.08f, 164.21f, 185.34f, 207.5f, 230.75f, 255.16f, 280.77f, 307.66f,
|
||||
335.9f, 365.56f, 396.71f, 429.44f, 463.84f, 500.f};
|
||||
const float kTestFilterBank[][2] = {{0.055556f, 0.f},
|
||||
{0.055556f, 0.f},
|
||||
{0.055556f, 0.f},
|
||||
{0.055556f, 0.f},
|
||||
{0.055556f, 0.f},
|
||||
{0.055556f, 0.f},
|
||||
{0.055556f, 0.f},
|
||||
{0.055556f, 0.f},
|
||||
{0.055556f, 0.f},
|
||||
{0.055556f, 0.f},
|
||||
{0.055556f, 0.f},
|
||||
{0.055556f, 0.f},
|
||||
{0.055556f, 0.f},
|
||||
{0.055556f, 0.f},
|
||||
{0.055556f, 0.f},
|
||||
{0.055556f, 0.f},
|
||||
{0.055556f, 0.f},
|
||||
{0.055556f, 0.2f},
|
||||
{0, 0.2f},
|
||||
{0, 0.2f},
|
||||
{0, 0.2f},
|
||||
{0, 0.2f}};
|
||||
const float kTestFilterBank[][9] = {
|
||||
{0.2f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f},
|
||||
{0.2f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f},
|
||||
{0.2f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f},
|
||||
{0.2f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f},
|
||||
{0.2f, 0.25f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f},
|
||||
{0.f, 0.25f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f},
|
||||
{0.f, 0.25f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f},
|
||||
{0.f, 0.25f, 0.25f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f},
|
||||
{0.f, 0.f, 0.25f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f},
|
||||
{0.f, 0.f, 0.25f, 0.142857f, 0.f, 0.f, 0.f, 0.f, 0.f},
|
||||
{0.f, 0.f, 0.25f, 0.285714f, 0.f, 0.f, 0.f, 0.f, 0.f},
|
||||
{0.f, 0.f, 0.f, 0.285714f, 0.142857f, 0.f, 0.f, 0.f, 0.f},
|
||||
{0.f, 0.f, 0.f, 0.285714f, 0.285714f, 0.f, 0.f, 0.f, 0.f},
|
||||
{0.f, 0.f, 0.f, 0.f, 0.285714f, 0.142857f, 0.f, 0.f, 0.f},
|
||||
{0.f, 0.f, 0.f, 0.f, 0.285714f, 0.285714f, 0.f, 0.f, 0.f},
|
||||
{0.f, 0.f, 0.f, 0.f, 0.f, 0.285714f, 0.142857f, 0.f, 0.f},
|
||||
{0.f, 0.f, 0.f, 0.f, 0.f, 0.285714f, 0.285714f, 0.f, 0.f},
|
||||
{0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.285714f, 0.142857f, 0.f},
|
||||
{0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.285714f, 0.285714f, 0.f},
|
||||
{0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.285714f, 0.f},
|
||||
{0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.285714f, 0.5f},
|
||||
{0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.5f}};
|
||||
static_assert(arraysize(kTestCenterFreqs) == arraysize(kTestFilterBank),
|
||||
"Test filterbank badly initialized.");
|
||||
|
||||
@ -63,14 +60,14 @@ 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, 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) == arraysize(kTestZeroVar),
|
||||
"Variance test data badly initialized.");
|
||||
"Power test data badly initialized.");
|
||||
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, 0.0351f, 0.0636f, 0.0863f,
|
||||
0.1037f, 0.1162f, 0.1236f, 0.1251f, 0.1189f, 0.0993f};
|
||||
static_assert(arraysize(kTestCenterFreqs) ==
|
||||
arraysize(kTestNonZeroVarLambdaTop),
|
||||
"Variance test data badly initialized.");
|
||||
"Power test data badly initialized.");
|
||||
const float kMaxTestError = 0.005f;
|
||||
|
||||
// Enhancer initialization parameters.
|
||||
@ -81,9 +78,6 @@ const int kFragmentSize = kSampleRate / 100;
|
||||
|
||||
} // namespace
|
||||
|
||||
using std::vector;
|
||||
using intelligibility::VarianceArray;
|
||||
|
||||
class IntelligibilityEnhancerTest : public ::testing::Test {
|
||||
protected:
|
||||
IntelligibilityEnhancerTest()
|
||||
@ -92,9 +86,8 @@ class IntelligibilityEnhancerTest : public ::testing::Test {
|
||||
enh_.reset(new IntelligibilityEnhancer(config_));
|
||||
}
|
||||
|
||||
bool CheckUpdate(VarianceArray::StepType step_type) {
|
||||
bool CheckUpdate() {
|
||||
config_.sample_rate_hz = kSampleRate;
|
||||
config_.var_type = step_type;
|
||||
enh_.reset(new IntelligibilityEnhancer(config_));
|
||||
float* clear_cursor = &clear_data_[0];
|
||||
float* noise_cursor = &noise_data_[0];
|
||||
@ -113,37 +106,25 @@ class IntelligibilityEnhancerTest : public ::testing::Test {
|
||||
|
||||
IntelligibilityEnhancer::Config config_;
|
||||
rtc::scoped_ptr<IntelligibilityEnhancer> enh_;
|
||||
vector<float> clear_data_;
|
||||
vector<float> noise_data_;
|
||||
vector<float> orig_data_;
|
||||
std::vector<float> clear_data_;
|
||||
std::vector<float> noise_data_;
|
||||
std::vector<float> orig_data_;
|
||||
};
|
||||
|
||||
// For each class of generated data, tests that render stream is
|
||||
// updated when it should be for each variance update method.
|
||||
// For each class of generated data, tests that render stream is updated when
|
||||
// it should be.
|
||||
TEST_F(IntelligibilityEnhancerTest, TestRenderUpdate) {
|
||||
vector<VarianceArray::StepType> step_types;
|
||||
step_types.push_back(VarianceArray::kStepInfinite);
|
||||
step_types.push_back(VarianceArray::kStepDecaying);
|
||||
step_types.push_back(VarianceArray::kStepWindowed);
|
||||
step_types.push_back(VarianceArray::kStepBlocked);
|
||||
step_types.push_back(VarianceArray::kStepBlockBasedMovingAverage);
|
||||
std::fill(noise_data_.begin(), noise_data_.end(), 0.0f);
|
||||
std::fill(orig_data_.begin(), orig_data_.end(), 0.0f);
|
||||
for (auto step_type : step_types) {
|
||||
std::fill(clear_data_.begin(), clear_data_.end(), 0.0f);
|
||||
EXPECT_FALSE(CheckUpdate(step_type));
|
||||
}
|
||||
std::fill(clear_data_.begin(), clear_data_.end(), 0.0f);
|
||||
EXPECT_FALSE(CheckUpdate());
|
||||
std::srand(1);
|
||||
auto float_rand = []() { return std::rand() * 2.f / RAND_MAX - 1; };
|
||||
std::generate(noise_data_.begin(), noise_data_.end(), float_rand);
|
||||
for (auto step_type : step_types) {
|
||||
EXPECT_FALSE(CheckUpdate(step_type));
|
||||
}
|
||||
for (auto step_type : step_types) {
|
||||
std::generate(clear_data_.begin(), clear_data_.end(), float_rand);
|
||||
orig_data_ = clear_data_;
|
||||
EXPECT_TRUE(CheckUpdate(step_type));
|
||||
}
|
||||
EXPECT_FALSE(CheckUpdate());
|
||||
std::generate(clear_data_.begin(), clear_data_.end(), float_rand);
|
||||
orig_data_ = clear_data_;
|
||||
EXPECT_TRUE(CheckUpdate());
|
||||
}
|
||||
|
||||
// Tests ERB bank creation, comparing against matlab output.
|
||||
@ -163,11 +144,11 @@ TEST_F(IntelligibilityEnhancerTest, TestErbCreation) {
|
||||
// against matlab output.
|
||||
TEST_F(IntelligibilityEnhancerTest, TestSolveForGains) {
|
||||
ASSERT_EQ(kTestStartFreq, enh_->start_freq_);
|
||||
vector<float> sols(enh_->bank_size_);
|
||||
std::vector<float> sols(enh_->bank_size_);
|
||||
float lambda = -0.001f;
|
||||
for (size_t i = 0; i < enh_->bank_size_; i++) {
|
||||
enh_->filtered_clear_var_[i] = 0.0f;
|
||||
enh_->filtered_noise_var_[i] = 0.0f;
|
||||
enh_->filtered_clear_pow_[i] = 0.0f;
|
||||
enh_->filtered_noise_pow_[i] = 0.0f;
|
||||
enh_->rho_[i] = 0.02f;
|
||||
}
|
||||
enh_->SolveForGainsGivenLambda(lambda, enh_->start_freq_, &sols[0]);
|
||||
@ -175,8 +156,8 @@ TEST_F(IntelligibilityEnhancerTest, TestSolveForGains) {
|
||||
EXPECT_NEAR(kTestZeroVar[i], sols[i], kMaxTestError);
|
||||
}
|
||||
for (size_t i = 0; i < enh_->bank_size_; i++) {
|
||||
enh_->filtered_clear_var_[i] = static_cast<float>(i + 1);
|
||||
enh_->filtered_noise_var_[i] = static_cast<float>(enh_->bank_size_ - i);
|
||||
enh_->filtered_clear_pow_[i] = static_cast<float>(i + 1);
|
||||
enh_->filtered_noise_pow_[i] = static_cast<float>(enh_->bank_size_ - i);
|
||||
}
|
||||
enh_->SolveForGainsGivenLambda(lambda, enh_->start_freq_, &sols[0]);
|
||||
for (size_t i = 0; i < enh_->bank_size_; i++) {
|
||||
|
||||
@ -8,10 +8,6 @@
|
||||
* be found in the AUTHORS file in the root of the source tree.
|
||||
*/
|
||||
|
||||
//
|
||||
// Implements helper functions and classes for intelligibility enhancement.
|
||||
//
|
||||
|
||||
#include "webrtc/modules/audio_processing/intelligibility/intelligibility_utils.h"
|
||||
|
||||
#include <math.h>
|
||||
@ -19,271 +15,46 @@
|
||||
#include <string.h>
|
||||
#include <algorithm>
|
||||
|
||||
using std::complex;
|
||||
using std::min;
|
||||
|
||||
namespace webrtc {
|
||||
|
||||
namespace intelligibility {
|
||||
|
||||
namespace {
|
||||
|
||||
// Return |current| changed towards |target|, with the change being at most
|
||||
// |limit|.
|
||||
float UpdateFactor(float target, float current, float limit) {
|
||||
float delta = fabsf(target - current);
|
||||
float sign = copysign(1.0f, target - current);
|
||||
float sign = copysign(1.f, target - current);
|
||||
return current + sign * fminf(delta, limit);
|
||||
}
|
||||
|
||||
float AddDitherIfZero(float value) {
|
||||
return value == 0.f ? std::rand() * 0.01f / RAND_MAX : value;
|
||||
}
|
||||
} // namespace
|
||||
|
||||
complex<float> zerofudge(complex<float> c) {
|
||||
return complex<float>(AddDitherIfZero(c.real()), AddDitherIfZero(c.imag()));
|
||||
}
|
||||
|
||||
complex<float> NewMean(complex<float> mean, complex<float> data, size_t count) {
|
||||
return mean + (data - mean) / static_cast<float>(count);
|
||||
}
|
||||
|
||||
void AddToMean(complex<float> data, size_t count, complex<float>* mean) {
|
||||
(*mean) = NewMean(*mean, data, count);
|
||||
}
|
||||
|
||||
|
||||
static const size_t kWindowBlockSize = 10;
|
||||
|
||||
VarianceArray::VarianceArray(size_t num_freqs,
|
||||
StepType type,
|
||||
size_t window_size,
|
||||
float decay)
|
||||
: running_mean_(new complex<float>[num_freqs]()),
|
||||
running_mean_sq_(new complex<float>[num_freqs]()),
|
||||
sub_running_mean_(new complex<float>[num_freqs]()),
|
||||
sub_running_mean_sq_(new complex<float>[num_freqs]()),
|
||||
variance_(new float[num_freqs]()),
|
||||
conj_sum_(new float[num_freqs]()),
|
||||
PowerEstimator::PowerEstimator(size_t num_freqs,
|
||||
float decay)
|
||||
: magnitude_(new float[num_freqs]()),
|
||||
power_(new float[num_freqs]()),
|
||||
num_freqs_(num_freqs),
|
||||
window_size_(window_size),
|
||||
decay_(decay),
|
||||
history_cursor_(0),
|
||||
count_(0),
|
||||
array_mean_(0.0f),
|
||||
buffer_full_(false) {
|
||||
history_.reset(new rtc::scoped_ptr<complex<float>[]>[num_freqs_]());
|
||||
for (size_t i = 0; i < num_freqs_; ++i) {
|
||||
history_[i].reset(new complex<float>[window_size_]());
|
||||
}
|
||||
subhistory_.reset(new rtc::scoped_ptr<complex<float>[]>[num_freqs_]());
|
||||
for (size_t i = 0; i < num_freqs_; ++i) {
|
||||
subhistory_[i].reset(new complex<float>[window_size_]());
|
||||
}
|
||||
subhistory_sq_.reset(new rtc::scoped_ptr<complex<float>[]>[num_freqs_]());
|
||||
for (size_t i = 0; i < num_freqs_; ++i) {
|
||||
subhistory_sq_[i].reset(new complex<float>[window_size_]());
|
||||
}
|
||||
switch (type) {
|
||||
case kStepInfinite:
|
||||
step_func_ = &VarianceArray::InfiniteStep;
|
||||
break;
|
||||
case kStepDecaying:
|
||||
step_func_ = &VarianceArray::DecayStep;
|
||||
break;
|
||||
case kStepWindowed:
|
||||
step_func_ = &VarianceArray::WindowedStep;
|
||||
break;
|
||||
case kStepBlocked:
|
||||
step_func_ = &VarianceArray::BlockedStep;
|
||||
break;
|
||||
case kStepBlockBasedMovingAverage:
|
||||
step_func_ = &VarianceArray::BlockBasedMovingAverage;
|
||||
break;
|
||||
}
|
||||
decay_(decay) {
|
||||
memset(magnitude_.get(), 0, sizeof(*magnitude_.get()) * num_freqs_);
|
||||
memset(power_.get(), 0, sizeof(*power_.get()) * num_freqs_);
|
||||
}
|
||||
|
||||
// Compute the variance with Welford's algorithm, adding some fudge to
|
||||
// the input in case of all-zeroes.
|
||||
void VarianceArray::InfiniteStep(const complex<float>* data, bool skip_fudge) {
|
||||
array_mean_ = 0.0f;
|
||||
++count_;
|
||||
for (size_t i = 0; i < num_freqs_; ++i) {
|
||||
complex<float> sample = data[i];
|
||||
if (!skip_fudge) {
|
||||
sample = zerofudge(sample);
|
||||
}
|
||||
if (count_ == 1) {
|
||||
running_mean_[i] = sample;
|
||||
variance_[i] = 0.0f;
|
||||
} else {
|
||||
float old_sum = conj_sum_[i];
|
||||
complex<float> old_mean = running_mean_[i];
|
||||
running_mean_[i] =
|
||||
old_mean + (sample - old_mean) / static_cast<float>(count_);
|
||||
conj_sum_[i] =
|
||||
(old_sum + std::conj(sample - old_mean) * (sample - running_mean_[i]))
|
||||
.real();
|
||||
variance_[i] =
|
||||
conj_sum_[i] / (count_ - 1);
|
||||
}
|
||||
array_mean_ += (variance_[i] - array_mean_) / (i + 1);
|
||||
}
|
||||
}
|
||||
|
||||
// Compute the variance from the beginning, with exponential decaying of the
|
||||
// Compute the magnitude from the beginning, with exponential decaying of the
|
||||
// series data.
|
||||
void VarianceArray::DecayStep(const complex<float>* data, bool /*dummy*/) {
|
||||
array_mean_ = 0.0f;
|
||||
++count_;
|
||||
void PowerEstimator::Step(const std::complex<float>* data) {
|
||||
for (size_t i = 0; i < num_freqs_; ++i) {
|
||||
complex<float> sample = data[i];
|
||||
sample = zerofudge(sample);
|
||||
|
||||
if (count_ == 1) {
|
||||
running_mean_[i] = sample;
|
||||
running_mean_sq_[i] = sample * std::conj(sample);
|
||||
variance_[i] = 0.0f;
|
||||
} else {
|
||||
complex<float> prev = running_mean_[i];
|
||||
complex<float> prev2 = running_mean_sq_[i];
|
||||
running_mean_[i] = decay_ * prev + (1.0f - decay_) * sample;
|
||||
running_mean_sq_[i] =
|
||||
decay_ * prev2 + (1.0f - decay_) * sample * std::conj(sample);
|
||||
variance_[i] = (running_mean_sq_[i] -
|
||||
running_mean_[i] * std::conj(running_mean_[i])).real();
|
||||
}
|
||||
|
||||
array_mean_ += (variance_[i] - array_mean_) / (i + 1);
|
||||
magnitude_[i] = decay_ * magnitude_[i] +
|
||||
(1.f - decay_) * std::abs(data[i]);
|
||||
}
|
||||
}
|
||||
|
||||
// Windowed variance computation. On each step, the variances for the
|
||||
// window are recomputed from scratch, using Welford's algorithm.
|
||||
void VarianceArray::WindowedStep(const complex<float>* data, bool /*dummy*/) {
|
||||
size_t num = min(count_ + 1, window_size_);
|
||||
array_mean_ = 0.0f;
|
||||
const float* PowerEstimator::Power() {
|
||||
for (size_t i = 0; i < num_freqs_; ++i) {
|
||||
complex<float> mean;
|
||||
float conj_sum = 0.0f;
|
||||
|
||||
history_[i][history_cursor_] = data[i];
|
||||
|
||||
mean = history_[i][history_cursor_];
|
||||
variance_[i] = 0.0f;
|
||||
for (size_t j = 1; j < num; ++j) {
|
||||
complex<float> sample =
|
||||
zerofudge(history_[i][(history_cursor_ + j) % window_size_]);
|
||||
sample = history_[i][(history_cursor_ + j) % window_size_];
|
||||
float old_sum = conj_sum;
|
||||
complex<float> old_mean = mean;
|
||||
|
||||
mean = old_mean + (sample - old_mean) / static_cast<float>(j + 1);
|
||||
conj_sum =
|
||||
(old_sum + std::conj(sample - old_mean) * (sample - mean)).real();
|
||||
variance_[i] = conj_sum / (j);
|
||||
}
|
||||
array_mean_ += (variance_[i] - array_mean_) / (i + 1);
|
||||
}
|
||||
history_cursor_ = (history_cursor_ + 1) % window_size_;
|
||||
++count_;
|
||||
}
|
||||
|
||||
// Variance with a window of blocks. Within each block, the variances are
|
||||
// recomputed from scratch at every stp, using |Var(X) = E(X^2) - E^2(X)|.
|
||||
// Once a block is filled with kWindowBlockSize samples, it is added to the
|
||||
// history window and a new block is started. The variances for the window
|
||||
// are recomputed from scratch at each of these transitions.
|
||||
void VarianceArray::BlockedStep(const complex<float>* data, bool /*dummy*/) {
|
||||
size_t blocks = min(window_size_, history_cursor_ + 1);
|
||||
for (size_t i = 0; i < num_freqs_; ++i) {
|
||||
AddToMean(data[i], count_ + 1, &sub_running_mean_[i]);
|
||||
AddToMean(data[i] * std::conj(data[i]), count_ + 1,
|
||||
&sub_running_mean_sq_[i]);
|
||||
subhistory_[i][history_cursor_ % window_size_] = sub_running_mean_[i];
|
||||
subhistory_sq_[i][history_cursor_ % window_size_] = sub_running_mean_sq_[i];
|
||||
|
||||
variance_[i] =
|
||||
(NewMean(running_mean_sq_[i], sub_running_mean_sq_[i], blocks) -
|
||||
NewMean(running_mean_[i], sub_running_mean_[i], blocks) *
|
||||
std::conj(NewMean(running_mean_[i], sub_running_mean_[i], blocks)))
|
||||
.real();
|
||||
if (count_ == kWindowBlockSize - 1) {
|
||||
sub_running_mean_[i] = complex<float>(0.0f, 0.0f);
|
||||
sub_running_mean_sq_[i] = complex<float>(0.0f, 0.0f);
|
||||
running_mean_[i] = complex<float>(0.0f, 0.0f);
|
||||
running_mean_sq_[i] = complex<float>(0.0f, 0.0f);
|
||||
for (size_t j = 0; j < min(window_size_, history_cursor_); ++j) {
|
||||
AddToMean(subhistory_[i][j], j + 1, &running_mean_[i]);
|
||||
AddToMean(subhistory_sq_[i][j], j + 1, &running_mean_sq_[i]);
|
||||
}
|
||||
++history_cursor_;
|
||||
}
|
||||
}
|
||||
++count_;
|
||||
if (count_ == kWindowBlockSize) {
|
||||
count_ = 0;
|
||||
}
|
||||
}
|
||||
|
||||
// Recomputes variances for each window from scratch based on previous window.
|
||||
void VarianceArray::BlockBasedMovingAverage(const std::complex<float>* data,
|
||||
bool /*dummy*/) {
|
||||
// TODO(ekmeyerson) To mitigate potential divergence, add counter so that
|
||||
// after every so often sums are computed scratch by summing over all
|
||||
// elements instead of subtracting oldest and adding newest.
|
||||
for (size_t i = 0; i < num_freqs_; ++i) {
|
||||
sub_running_mean_[i] += data[i];
|
||||
sub_running_mean_sq_[i] += data[i] * std::conj(data[i]);
|
||||
}
|
||||
++count_;
|
||||
|
||||
// TODO(ekmeyerson) Make kWindowBlockSize nonconstant to allow
|
||||
// experimentation with different block size,window size pairs.
|
||||
if (count_ >= kWindowBlockSize) {
|
||||
count_ = 0;
|
||||
|
||||
for (size_t i = 0; i < num_freqs_; ++i) {
|
||||
running_mean_[i] -= subhistory_[i][history_cursor_];
|
||||
running_mean_sq_[i] -= subhistory_sq_[i][history_cursor_];
|
||||
|
||||
float scale = 1.f / kWindowBlockSize;
|
||||
subhistory_[i][history_cursor_] = sub_running_mean_[i] * scale;
|
||||
subhistory_sq_[i][history_cursor_] = sub_running_mean_sq_[i] * scale;
|
||||
|
||||
sub_running_mean_[i] = std::complex<float>(0.0f, 0.0f);
|
||||
sub_running_mean_sq_[i] = std::complex<float>(0.0f, 0.0f);
|
||||
|
||||
running_mean_[i] += subhistory_[i][history_cursor_];
|
||||
running_mean_sq_[i] += subhistory_sq_[i][history_cursor_];
|
||||
|
||||
scale = 1.f / (buffer_full_ ? window_size_ : history_cursor_ + 1);
|
||||
variance_[i] = std::real(running_mean_sq_[i] * scale -
|
||||
running_mean_[i] * scale *
|
||||
std::conj(running_mean_[i]) * scale);
|
||||
}
|
||||
|
||||
++history_cursor_;
|
||||
if (history_cursor_ >= window_size_) {
|
||||
buffer_full_ = true;
|
||||
history_cursor_ = 0;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
void VarianceArray::Clear() {
|
||||
memset(running_mean_.get(), 0, sizeof(*running_mean_.get()) * num_freqs_);
|
||||
memset(running_mean_sq_.get(), 0,
|
||||
sizeof(*running_mean_sq_.get()) * num_freqs_);
|
||||
memset(variance_.get(), 0, sizeof(*variance_.get()) * num_freqs_);
|
||||
memset(conj_sum_.get(), 0, sizeof(*conj_sum_.get()) * num_freqs_);
|
||||
history_cursor_ = 0;
|
||||
count_ = 0;
|
||||
array_mean_ = 0.0f;
|
||||
}
|
||||
|
||||
void VarianceArray::ApplyScale(float scale) {
|
||||
array_mean_ = 0.0f;
|
||||
for (size_t i = 0; i < num_freqs_; ++i) {
|
||||
variance_[i] *= scale * scale;
|
||||
array_mean_ += (variance_[i] - array_mean_) / (i + 1);
|
||||
power_[i] = magnitude_[i] * magnitude_[i];
|
||||
}
|
||||
return &power_[0];
|
||||
}
|
||||
|
||||
GainApplier::GainApplier(size_t freqs, float change_limit)
|
||||
@ -292,17 +63,17 @@ GainApplier::GainApplier(size_t freqs, float change_limit)
|
||||
target_(new float[freqs]()),
|
||||
current_(new float[freqs]()) {
|
||||
for (size_t i = 0; i < freqs; ++i) {
|
||||
target_[i] = 1.0f;
|
||||
current_[i] = 1.0f;
|
||||
target_[i] = 1.f;
|
||||
current_[i] = 1.f;
|
||||
}
|
||||
}
|
||||
|
||||
void GainApplier::Apply(const complex<float>* in_block,
|
||||
complex<float>* out_block) {
|
||||
void GainApplier::Apply(const std::complex<float>* in_block,
|
||||
std::complex<float>* out_block) {
|
||||
for (size_t i = 0; i < num_freqs_; ++i) {
|
||||
float factor = sqrtf(fabsf(current_[i]));
|
||||
if (!std::isnormal(factor)) {
|
||||
factor = 1.0f;
|
||||
factor = 1.f;
|
||||
}
|
||||
out_block[i] = factor * in_block[i];
|
||||
current_[i] = UpdateFactor(target_[i], current_[i], change_limit_);
|
||||
|
||||
@ -8,10 +8,6 @@
|
||||
* be found in the AUTHORS file in the root of the source tree.
|
||||
*/
|
||||
|
||||
//
|
||||
// Specifies helper classes for intelligibility enhancement.
|
||||
//
|
||||
|
||||
#ifndef WEBRTC_MODULES_AUDIO_PROCESSING_INTELLIGIBILITY_INTELLIGIBILITY_UTILS_H_
|
||||
#define WEBRTC_MODULES_AUDIO_PROCESSING_INTELLIGIBILITY_INTELLIGIBILITY_UTILS_H_
|
||||
|
||||
@ -23,115 +19,36 @@ namespace webrtc {
|
||||
|
||||
namespace intelligibility {
|
||||
|
||||
// Return |current| changed towards |target|, with the change being at most
|
||||
// |limit|.
|
||||
float UpdateFactor(float target, float current, float limit);
|
||||
|
||||
// Apply a small fudge to degenerate complex values. The numbers in the array
|
||||
// were chosen randomly, so that even a series of all zeroes has some small
|
||||
// variability.
|
||||
std::complex<float> zerofudge(std::complex<float> c);
|
||||
|
||||
// Incremental mean computation. Return the mean of the series with the
|
||||
// mean |mean| with added |data|.
|
||||
std::complex<float> NewMean(std::complex<float> mean,
|
||||
std::complex<float> data,
|
||||
size_t count);
|
||||
|
||||
// Updates |mean| with added |data|;
|
||||
void AddToMean(std::complex<float> data,
|
||||
size_t count,
|
||||
std::complex<float>* mean);
|
||||
|
||||
// Internal helper for computing the variances of a stream of arrays.
|
||||
// The result is an array of variances per position: the i-th variance
|
||||
// is the variance of the stream of data on the i-th positions in the
|
||||
// input arrays.
|
||||
// There are four methods of computation:
|
||||
// * kStepInfinite computes variances from the beginning onwards
|
||||
// * kStepDecaying uses a recursive exponential decay formula with a
|
||||
// settable forgetting factor
|
||||
// * kStepWindowed computes variances within a moving window
|
||||
// * kStepBlocked is similar to kStepWindowed, but history is kept
|
||||
// as a rolling window of blocks: multiple input elements are used for
|
||||
// one block and the history then consists of the variances of these blocks
|
||||
// with the same effect as kStepWindowed, but less storage, so the window
|
||||
// can be longer
|
||||
class VarianceArray {
|
||||
// Internal helper for computing the power of a stream of arrays.
|
||||
// The result is an array of power per position: the i-th power is the power of
|
||||
// the stream of data on the i-th positions in the input arrays.
|
||||
class PowerEstimator {
|
||||
public:
|
||||
enum StepType {
|
||||
kStepInfinite = 0,
|
||||
kStepDecaying,
|
||||
kStepWindowed,
|
||||
kStepBlocked,
|
||||
kStepBlockBasedMovingAverage
|
||||
};
|
||||
// Construct an instance for the given input array length (|freqs|), with the
|
||||
// appropriate parameters. |decay| is the forgetting factor.
|
||||
PowerEstimator(size_t freqs, float decay);
|
||||
|
||||
// Construct an instance for the given input array length (|freqs|) and
|
||||
// computation algorithm (|type|), with the appropriate parameters.
|
||||
// |window_size| is the number of samples for kStepWindowed and
|
||||
// the number of blocks for kStepBlocked. |decay| is the forgetting factor
|
||||
// for kStepDecaying.
|
||||
VarianceArray(size_t freqs, StepType type, size_t window_size, float decay);
|
||||
// Add a new data point to the series.
|
||||
void Step(const std::complex<float>* data);
|
||||
|
||||
// Add a new data point to the series and compute the new variances.
|
||||
// TODO(bercic) |skip_fudge| is a flag for kStepWindowed and kStepDecaying,
|
||||
// whether they should skip adding some small dummy values to the input
|
||||
// to prevent problems with all-zero inputs. Can probably be removed.
|
||||
void Step(const std::complex<float>* data, bool skip_fudge = false) {
|
||||
(this->*step_func_)(data, skip_fudge);
|
||||
}
|
||||
// Reset variances to zero and forget all history.
|
||||
void Clear();
|
||||
// Scale the input data by |scale|. Effectively multiply variances
|
||||
// by |scale^2|.
|
||||
void ApplyScale(float scale);
|
||||
|
||||
// The current set of variances.
|
||||
const float* variance() const { return variance_.get(); }
|
||||
|
||||
// The mean value of the current set of variances.
|
||||
float array_mean() const { return array_mean_; }
|
||||
// The current power array.
|
||||
const float* Power();
|
||||
|
||||
private:
|
||||
void InfiniteStep(const std::complex<float>* data, bool dummy);
|
||||
void DecayStep(const std::complex<float>* data, bool dummy);
|
||||
void WindowedStep(const std::complex<float>* data, bool dummy);
|
||||
void BlockedStep(const std::complex<float>* data, bool dummy);
|
||||
void BlockBasedMovingAverage(const std::complex<float>* data, bool dummy);
|
||||
|
||||
// TODO(ekmeyerson): Switch the following running means
|
||||
// and histories from rtc::scoped_ptr to std::vector.
|
||||
|
||||
// The current average X and X^2.
|
||||
rtc::scoped_ptr<std::complex<float>[]> running_mean_;
|
||||
rtc::scoped_ptr<std::complex<float>[]> running_mean_sq_;
|
||||
|
||||
// Average X and X^2 for the current block in kStepBlocked.
|
||||
rtc::scoped_ptr<std::complex<float>[]> sub_running_mean_;
|
||||
rtc::scoped_ptr<std::complex<float>[]> sub_running_mean_sq_;
|
||||
|
||||
// Sample history for the rolling window in kStepWindowed and block-wise
|
||||
// histories for kStepBlocked.
|
||||
rtc::scoped_ptr<rtc::scoped_ptr<std::complex<float>[]>[]> history_;
|
||||
rtc::scoped_ptr<rtc::scoped_ptr<std::complex<float>[]>[]> subhistory_;
|
||||
rtc::scoped_ptr<rtc::scoped_ptr<std::complex<float>[]>[]> subhistory_sq_;
|
||||
|
||||
// The current set of variances and sums for Welford's algorithm.
|
||||
rtc::scoped_ptr<float[]> variance_;
|
||||
rtc::scoped_ptr<float[]> conj_sum_;
|
||||
// The current magnitude array.
|
||||
rtc::scoped_ptr<float[]> magnitude_;
|
||||
// The current power array.
|
||||
rtc::scoped_ptr<float[]> power_;
|
||||
|
||||
const size_t num_freqs_;
|
||||
const size_t window_size_;
|
||||
const float decay_;
|
||||
size_t history_cursor_;
|
||||
size_t count_;
|
||||
float array_mean_;
|
||||
bool buffer_full_;
|
||||
void (VarianceArray::*step_func_)(const std::complex<float>*, bool);
|
||||
};
|
||||
|
||||
// Helper class for smoothing gain changes. On each applicatiion step, the
|
||||
// Helper class for smoothing gain changes. On each application step, the
|
||||
// currently used gains are changed towards a set of settable target gains,
|
||||
// constrained by a limit on the magnitude of the changes.
|
||||
class GainApplier {
|
||||
|
||||
@ -8,169 +8,69 @@
|
||||
* be found in the AUTHORS file in the root of the source tree.
|
||||
*/
|
||||
|
||||
//
|
||||
// Unit tests for intelligibility utils.
|
||||
//
|
||||
|
||||
#include <math.h>
|
||||
#include <cmath>
|
||||
#include <complex>
|
||||
#include <iostream>
|
||||
#include <vector>
|
||||
|
||||
#include "testing/gtest/include/gtest/gtest.h"
|
||||
#include "webrtc/base/arraysize.h"
|
||||
#include "webrtc/modules/audio_processing/intelligibility/intelligibility_utils.h"
|
||||
|
||||
using std::complex;
|
||||
using std::vector;
|
||||
|
||||
namespace webrtc {
|
||||
|
||||
namespace intelligibility {
|
||||
|
||||
vector<vector<complex<float>>> GenerateTestData(int freqs, int samples) {
|
||||
vector<vector<complex<float>>> data(samples);
|
||||
for (int i = 0; i < samples; i++) {
|
||||
for (int j = 0; j < freqs; j++) {
|
||||
std::vector<std::vector<std::complex<float>>> GenerateTestData(size_t freqs,
|
||||
size_t samples) {
|
||||
std::vector<std::vector<std::complex<float>>> data(samples);
|
||||
for (size_t i = 0; i < samples; ++i) {
|
||||
for (size_t j = 0; j < freqs; ++j) {
|
||||
const float val = 0.99f / ((i + 1) * (j + 1));
|
||||
data[i].push_back(complex<float>(val, val));
|
||||
data[i].push_back(std::complex<float>(val, val));
|
||||
}
|
||||
}
|
||||
return data;
|
||||
}
|
||||
|
||||
// Tests UpdateFactor.
|
||||
TEST(IntelligibilityUtilsTest, TestUpdateFactor) {
|
||||
EXPECT_EQ(0, intelligibility::UpdateFactor(0, 0, 0));
|
||||
EXPECT_EQ(4, intelligibility::UpdateFactor(4, 2, 3));
|
||||
EXPECT_EQ(3, intelligibility::UpdateFactor(4, 2, 1));
|
||||
EXPECT_EQ(2, intelligibility::UpdateFactor(2, 4, 3));
|
||||
EXPECT_EQ(3, intelligibility::UpdateFactor(2, 4, 1));
|
||||
}
|
||||
|
||||
// Tests zerofudge.
|
||||
TEST(IntelligibilityUtilsTest, TestCplx) {
|
||||
complex<float> t0(1.f, 0.f);
|
||||
t0 = intelligibility::zerofudge(t0);
|
||||
EXPECT_NE(t0.imag(), 0.f);
|
||||
EXPECT_NE(t0.real(), 0.f);
|
||||
}
|
||||
|
||||
// Tests NewMean and AddToMean.
|
||||
TEST(IntelligibilityUtilsTest, TestMeanUpdate) {
|
||||
const complex<float> data[] = {{3, 8}, {7, 6}, {2, 1}, {8, 9}, {0, 6}};
|
||||
const complex<float> means[] = {{3, 8}, {5, 7}, {4, 5}, {5, 6}, {4, 6}};
|
||||
complex<float> mean(3, 8);
|
||||
for (size_t i = 0; i < arraysize(data); i++) {
|
||||
EXPECT_EQ(means[i], NewMean(mean, data[i], i + 1));
|
||||
AddToMean(data[i], i + 1, &mean);
|
||||
EXPECT_EQ(means[i], mean);
|
||||
}
|
||||
}
|
||||
|
||||
// Tests VarianceArray, for all variance step types.
|
||||
TEST(IntelligibilityUtilsTest, TestVarianceArray) {
|
||||
const int kFreqs = 10;
|
||||
const int kSamples = 100;
|
||||
const int kWindowSize = 10; // Should pass for all kWindowSize > 1.
|
||||
// Tests PowerEstimator, for all power step types.
|
||||
TEST(IntelligibilityUtilsTest, TestPowerEstimator) {
|
||||
const size_t kFreqs = 10;
|
||||
const size_t kSamples = 100;
|
||||
const float kDecay = 0.5f;
|
||||
vector<VarianceArray::StepType> step_types;
|
||||
step_types.push_back(VarianceArray::kStepInfinite);
|
||||
step_types.push_back(VarianceArray::kStepDecaying);
|
||||
step_types.push_back(VarianceArray::kStepWindowed);
|
||||
step_types.push_back(VarianceArray::kStepBlocked);
|
||||
step_types.push_back(VarianceArray::kStepBlockBasedMovingAverage);
|
||||
const vector<vector<complex<float>>> test_data(
|
||||
const std::vector<std::vector<std::complex<float>>> test_data(
|
||||
GenerateTestData(kFreqs, kSamples));
|
||||
for (auto step_type : step_types) {
|
||||
VarianceArray variance_array(kFreqs, step_type, kWindowSize, kDecay);
|
||||
EXPECT_EQ(0, variance_array.variance()[0]);
|
||||
EXPECT_EQ(0, variance_array.array_mean());
|
||||
variance_array.ApplyScale(2.0f);
|
||||
EXPECT_EQ(0, variance_array.variance()[0]);
|
||||
EXPECT_EQ(0, variance_array.array_mean());
|
||||
PowerEstimator power_estimator(kFreqs, kDecay);
|
||||
EXPECT_EQ(0, power_estimator.Power()[0]);
|
||||
|
||||
// Makes sure Step is doing something.
|
||||
variance_array.Step(&test_data[0][0]);
|
||||
for (int i = 1; i < kSamples; i++) {
|
||||
variance_array.Step(&test_data[i][0]);
|
||||
EXPECT_GE(variance_array.array_mean(), 0.0f);
|
||||
EXPECT_LE(variance_array.array_mean(), 1.0f);
|
||||
for (int j = 0; j < kFreqs; j++) {
|
||||
EXPECT_GE(variance_array.variance()[j], 0.0f);
|
||||
EXPECT_LE(variance_array.variance()[j], 1.0f);
|
||||
}
|
||||
}
|
||||
variance_array.Clear();
|
||||
EXPECT_EQ(0, variance_array.variance()[0]);
|
||||
EXPECT_EQ(0, variance_array.array_mean());
|
||||
}
|
||||
}
|
||||
|
||||
// Tests exact computation on synthetic data.
|
||||
TEST(IntelligibilityUtilsTest, TestMovingBlockAverage) {
|
||||
// Exact, not unbiased estimates.
|
||||
const float kTestVarianceBufferNotFull = 16.5f;
|
||||
const float kTestVarianceBufferFull1 = 66.5f;
|
||||
const float kTestVarianceBufferFull2 = 333.375f;
|
||||
const int kFreqs = 2;
|
||||
const int kSamples = 50;
|
||||
const int kWindowSize = 2;
|
||||
const float kDecay = 0.5f;
|
||||
const float kMaxError = 0.0001f;
|
||||
|
||||
VarianceArray variance_array(
|
||||
kFreqs, VarianceArray::kStepBlockBasedMovingAverage, kWindowSize, kDecay);
|
||||
|
||||
vector<vector<complex<float>>> test_data(kSamples);
|
||||
for (int i = 0; i < kSamples; i++) {
|
||||
for (int j = 0; j < kFreqs; j++) {
|
||||
if (i < 30) {
|
||||
test_data[i].push_back(complex<float>(static_cast<float>(kSamples - i),
|
||||
static_cast<float>(i + 1)));
|
||||
} else {
|
||||
test_data[i].push_back(complex<float>(0.f, 0.f));
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
for (int i = 0; i < kSamples; i++) {
|
||||
variance_array.Step(&test_data[i][0]);
|
||||
for (int j = 0; j < kFreqs; j++) {
|
||||
if (i < 9) { // In utils, kWindowBlockSize = 10.
|
||||
EXPECT_EQ(0, variance_array.variance()[j]);
|
||||
} else if (i < 19) {
|
||||
EXPECT_NEAR(kTestVarianceBufferNotFull, variance_array.variance()[j],
|
||||
kMaxError);
|
||||
} else if (i < 39) {
|
||||
EXPECT_NEAR(kTestVarianceBufferFull1, variance_array.variance()[j],
|
||||
kMaxError);
|
||||
} else if (i < 49) {
|
||||
EXPECT_NEAR(kTestVarianceBufferFull2, variance_array.variance()[j],
|
||||
kMaxError);
|
||||
} else {
|
||||
EXPECT_EQ(0, variance_array.variance()[j]);
|
||||
}
|
||||
// Makes sure Step is doing something.
|
||||
power_estimator.Step(&test_data[0][0]);
|
||||
for (size_t i = 1; i < kSamples; ++i) {
|
||||
power_estimator.Step(&test_data[i][0]);
|
||||
for (size_t j = 0; j < kFreqs; ++j) {
|
||||
const float* power = power_estimator.Power();
|
||||
EXPECT_GE(power[j], 0.f);
|
||||
EXPECT_LE(power[j], 1.f);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// Tests gain applier.
|
||||
TEST(IntelligibilityUtilsTest, TestGainApplier) {
|
||||
const int kFreqs = 10;
|
||||
const int kSamples = 100;
|
||||
const size_t kFreqs = 10;
|
||||
const size_t kSamples = 100;
|
||||
const float kChangeLimit = 0.1f;
|
||||
GainApplier gain_applier(kFreqs, kChangeLimit);
|
||||
const vector<vector<complex<float>>> in_data(
|
||||
const std::vector<std::vector<std::complex<float>>> in_data(
|
||||
GenerateTestData(kFreqs, kSamples));
|
||||
vector<vector<complex<float>>> out_data(GenerateTestData(kFreqs, kSamples));
|
||||
for (int i = 0; i < kSamples; i++) {
|
||||
std::vector<std::vector<std::complex<float>>> out_data(GenerateTestData(
|
||||
kFreqs, kSamples));
|
||||
for (size_t i = 0; i < kSamples; ++i) {
|
||||
gain_applier.Apply(&in_data[i][0], &out_data[i][0]);
|
||||
for (int j = 0; j < kFreqs; j++) {
|
||||
EXPECT_GT(out_data[i][j].real(), 0.0f);
|
||||
EXPECT_LT(out_data[i][j].real(), 1.0f);
|
||||
EXPECT_GT(out_data[i][j].imag(), 0.0f);
|
||||
EXPECT_LT(out_data[i][j].imag(), 1.0f);
|
||||
for (size_t j = 0; j < kFreqs; ++j) {
|
||||
EXPECT_GT(out_data[i][j].real(), 0.f);
|
||||
EXPECT_LT(out_data[i][j].real(), 1.f);
|
||||
EXPECT_GT(out_data[i][j].imag(), 0.f);
|
||||
EXPECT_LT(out_data[i][j].imag(), 1.f);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
@ -24,6 +24,7 @@
|
||||
#include "testing/gtest/include/gtest/gtest.h"
|
||||
#include "webrtc/base/checks.h"
|
||||
#include "webrtc/base/criticalsection.h"
|
||||
#include "webrtc/common_audio/include/audio_util.h"
|
||||
#include "webrtc/common_audio/real_fourier.h"
|
||||
#include "webrtc/common_audio/wav_file.h"
|
||||
#include "webrtc/modules/audio_processing/audio_buffer.h"
|
||||
@ -35,34 +36,17 @@
|
||||
#include "webrtc/test/testsupport/fileutils.h"
|
||||
|
||||
using std::complex;
|
||||
using webrtc::intelligibility::VarianceArray;
|
||||
|
||||
namespace webrtc {
|
||||
namespace {
|
||||
|
||||
bool ValidateClearWindow(const char* flagname, int32_t value) {
|
||||
return value > 0;
|
||||
}
|
||||
|
||||
DEFINE_int32(clear_type,
|
||||
webrtc::intelligibility::VarianceArray::kStepDecaying,
|
||||
"Variance algorithm for clear data.");
|
||||
DEFINE_double(clear_alpha, 0.9, "Variance decay factor for clear data.");
|
||||
DEFINE_int32(clear_window,
|
||||
475,
|
||||
"Window size for windowed variance for clear data.");
|
||||
const bool clear_window_dummy =
|
||||
google::RegisterFlagValidator(&FLAGS_clear_window, &ValidateClearWindow);
|
||||
DEFINE_double(clear_alpha, 0.9, "Power decay factor for clear data.");
|
||||
DEFINE_int32(sample_rate,
|
||||
16000,
|
||||
"Audio sample rate used in the input and output files.");
|
||||
DEFINE_int32(ana_rate,
|
||||
800,
|
||||
60,
|
||||
"Analysis rate; gains recalculated every N blocks.");
|
||||
DEFINE_int32(
|
||||
var_rate,
|
||||
2,
|
||||
"Variance clear rate; history is forgotten every N gain recalculations.");
|
||||
DEFINE_double(gain_limit, 1000.0, "Maximum gain change in one block.");
|
||||
|
||||
DEFINE_string(clear_file, "speech.wav", "Input file with clear speech.");
|
||||
@ -77,11 +61,7 @@ const size_t kNumChannels = 1;
|
||||
// void function for gtest
|
||||
void void_main(int argc, char* argv[]) {
|
||||
google::SetUsageMessage(
|
||||
"\n\nVariance algorithm types are:\n"
|
||||
" 0 - infinite/normal,\n"
|
||||
" 1 - exponentially decaying,\n"
|
||||
" 2 - rolling window.\n"
|
||||
"\nInput files must be little-endian 16-bit signed raw PCM.\n");
|
||||
"\n\nInput files must be little-endian 16-bit signed raw PCM.\n");
|
||||
google::ParseCommandLineFlags(&argc, &argv, true);
|
||||
|
||||
size_t samples; // Number of samples in input PCM file
|
||||
@ -105,17 +85,17 @@ void void_main(int argc, char* argv[]) {
|
||||
WavReader in_file(FLAGS_clear_file);
|
||||
std::vector<float> in_fpcm(samples);
|
||||
in_file.ReadSamples(samples, &in_fpcm[0]);
|
||||
FloatS16ToFloat(&in_fpcm[0], samples, &in_fpcm[0]);
|
||||
|
||||
WavReader noise_file(FLAGS_noise_file);
|
||||
std::vector<float> noise_fpcm(samples);
|
||||
noise_file.ReadSamples(samples, &noise_fpcm[0]);
|
||||
FloatS16ToFloat(&noise_fpcm[0], samples, &noise_fpcm[0]);
|
||||
|
||||
// Run intelligibility enhancement.
|
||||
IntelligibilityEnhancer::Config config;
|
||||
config.sample_rate_hz = FLAGS_sample_rate;
|
||||
config.var_type = static_cast<VarianceArray::StepType>(FLAGS_clear_type);
|
||||
config.var_decay_rate = static_cast<float>(FLAGS_clear_alpha);
|
||||
config.var_window_size = static_cast<size_t>(FLAGS_clear_window);
|
||||
config.decay_rate = static_cast<float>(FLAGS_clear_alpha);
|
||||
config.analysis_rate = FLAGS_ana_rate;
|
||||
config.gain_change_limit = FLAGS_gain_limit;
|
||||
IntelligibilityEnhancer enh(config);
|
||||
@ -146,6 +126,8 @@ void void_main(int argc, char* argv[]) {
|
||||
noise_cursor += fragment_size;
|
||||
}
|
||||
|
||||
FloatToFloatS16(&in_fpcm[0], samples, &in_fpcm[0]);
|
||||
|
||||
if (FLAGS_out_file.compare("-") == 0) {
|
||||
const std::string temp_out_filename =
|
||||
test::TempFilename(test::WorkingDir(), "temp_wav_file");
|
||||
|
||||
@ -177,15 +177,17 @@ std::vector<float> NoiseSuppressionImpl::NoiseEstimate() {
|
||||
rtc::CritScope cs(crit_);
|
||||
std::vector<float> noise_estimate;
|
||||
#if defined(WEBRTC_NS_FLOAT)
|
||||
const float kNormalizationFactor = 1.f / (1 << 15);
|
||||
noise_estimate.assign(WebRtcNs_num_freq(), 0.f);
|
||||
for (auto& suppressor : suppressors_) {
|
||||
const float* noise = WebRtcNs_noise_estimate(suppressor->state());
|
||||
for (size_t i = 0; i < noise_estimate.size(); ++i) {
|
||||
noise_estimate[i] += noise[i] / suppressors_.size();
|
||||
noise_estimate[i] += kNormalizationFactor *
|
||||
noise[i] / suppressors_.size();
|
||||
}
|
||||
}
|
||||
#elif defined(WEBRTC_NS_FIXED)
|
||||
const float kNormalizationFactor = 1.f / (1 << 8);
|
||||
const float kNormalizationFactor = 1.f / (1 << 23);
|
||||
noise_estimate.assign(WebRtcNsx_num_freq(), 0.f);
|
||||
for (auto& suppressor : suppressors_) {
|
||||
const uint32_t* noise = WebRtcNsx_noise_estimate(suppressor->state());
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user