1
0
mirror of https://github.com/danog/libtgvoip.git synced 2024-12-11 16:49:52 +01:00
libtgvoip/webrtc_dsp/modules/audio_processing/transient/transient_detector.cc
Grishka 5caaaafa42 Updated WebRTC APM
I'm now using the entire audio processing module from WebRTC as opposed to individual DSP algorithms pulled from there before. Seems to work better this way.
2018-11-23 04:02:53 +03:00

176 lines
6.3 KiB
C++

/*
* Copyright (c) 2013 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/transient/transient_detector.h"
#include <float.h>
#include <math.h>
#include <string.h>
#include <algorithm>
#include "modules/audio_processing/transient/common.h"
#include "modules/audio_processing/transient/daubechies_8_wavelet_coeffs.h"
#include "modules/audio_processing/transient/moving_moments.h"
#include "modules/audio_processing/transient/wpd_node.h"
#include "modules/audio_processing/transient/wpd_tree.h"
#include "rtc_base/checks.h"
namespace webrtc {
static const int kTransientLengthMs = 30;
static const int kChunksAtStartupLeftToDelete =
kTransientLengthMs / ts::kChunkSizeMs;
static const float kDetectThreshold = 16.f;
TransientDetector::TransientDetector(int sample_rate_hz)
: samples_per_chunk_(sample_rate_hz * ts::kChunkSizeMs / 1000),
last_first_moment_(),
last_second_moment_(),
chunks_at_startup_left_to_delete_(kChunksAtStartupLeftToDelete),
reference_energy_(1.f),
using_reference_(false) {
RTC_DCHECK(sample_rate_hz == ts::kSampleRate8kHz ||
sample_rate_hz == ts::kSampleRate16kHz ||
sample_rate_hz == ts::kSampleRate32kHz ||
sample_rate_hz == ts::kSampleRate48kHz);
int samples_per_transient = sample_rate_hz * kTransientLengthMs / 1000;
// Adjustment to avoid data loss while downsampling, making
// |samples_per_chunk_| and |samples_per_transient| always divisible by
// |kLeaves|.
samples_per_chunk_ -= samples_per_chunk_ % kLeaves;
samples_per_transient -= samples_per_transient % kLeaves;
tree_leaves_data_length_ = samples_per_chunk_ / kLeaves;
wpd_tree_.reset(new WPDTree(samples_per_chunk_,
kDaubechies8HighPassCoefficients,
kDaubechies8LowPassCoefficients,
kDaubechies8CoefficientsLength, kLevels));
for (size_t i = 0; i < kLeaves; ++i) {
moving_moments_[i].reset(
new MovingMoments(samples_per_transient / kLeaves));
}
first_moments_.reset(new float[tree_leaves_data_length_]);
second_moments_.reset(new float[tree_leaves_data_length_]);
for (int i = 0; i < kChunksAtStartupLeftToDelete; ++i) {
previous_results_.push_back(0.f);
}
}
TransientDetector::~TransientDetector() {}
float TransientDetector::Detect(const float* data,
size_t data_length,
const float* reference_data,
size_t reference_length) {
RTC_DCHECK(data);
RTC_DCHECK_EQ(samples_per_chunk_, data_length);
// TODO(aluebs): Check if these errors can logically happen and if not assert
// on them.
if (wpd_tree_->Update(data, samples_per_chunk_) != 0) {
return -1.f;
}
float result = 0.f;
for (size_t i = 0; i < kLeaves; ++i) {
WPDNode* leaf = wpd_tree_->NodeAt(kLevels, i);
moving_moments_[i]->CalculateMoments(leaf->data(), tree_leaves_data_length_,
first_moments_.get(),
second_moments_.get());
// Add value delayed (Use the last moments from the last call to Detect).
float unbiased_data = leaf->data()[0] - last_first_moment_[i];
result +=
unbiased_data * unbiased_data / (last_second_moment_[i] + FLT_MIN);
// Add new values.
for (size_t j = 1; j < tree_leaves_data_length_; ++j) {
unbiased_data = leaf->data()[j] - first_moments_[j - 1];
result +=
unbiased_data * unbiased_data / (second_moments_[j - 1] + FLT_MIN);
}
last_first_moment_[i] = first_moments_[tree_leaves_data_length_ - 1];
last_second_moment_[i] = second_moments_[tree_leaves_data_length_ - 1];
}
result /= tree_leaves_data_length_;
result *= ReferenceDetectionValue(reference_data, reference_length);
if (chunks_at_startup_left_to_delete_ > 0) {
chunks_at_startup_left_to_delete_--;
result = 0.f;
}
if (result >= kDetectThreshold) {
result = 1.f;
} else {
// Get proportional value.
// Proportion achieved with a squared raised cosine function with domain
// [0, kDetectThreshold) and image [0, 1), it's always increasing.
const float horizontal_scaling = ts::kPi / kDetectThreshold;
const float kHorizontalShift = ts::kPi;
const float kVerticalScaling = 0.5f;
const float kVerticalShift = 1.f;
result =
(cos(result * horizontal_scaling + kHorizontalShift) + kVerticalShift) *
kVerticalScaling;
result *= result;
}
previous_results_.pop_front();
previous_results_.push_back(result);
// In the current implementation we return the max of the current result and
// the previous results, so the high results have a width equals to
// |transient_length|.
return *std::max_element(previous_results_.begin(), previous_results_.end());
}
// Looks for the highest slope and compares it with the previous ones.
// An exponential transformation takes this to the [0, 1] range. This value is
// multiplied by the detection result to avoid false positives.
float TransientDetector::ReferenceDetectionValue(const float* data,
size_t length) {
if (data == NULL) {
using_reference_ = false;
return 1.f;
}
static const float kEnergyRatioThreshold = 0.2f;
static const float kReferenceNonLinearity = 20.f;
static const float kMemory = 0.99f;
float reference_energy = 0.f;
for (size_t i = 1; i < length; ++i) {
reference_energy += data[i] * data[i];
}
if (reference_energy == 0.f) {
using_reference_ = false;
return 1.f;
}
RTC_DCHECK_NE(0, reference_energy_);
float result = 1.f / (1.f + exp(kReferenceNonLinearity *
(kEnergyRatioThreshold -
reference_energy / reference_energy_)));
reference_energy_ =
kMemory * reference_energy_ + (1.f - kMemory) * reference_energy;
using_reference_ = true;
return result;
}
} // namespace webrtc