d94659dc27
This is the first public upload of the new NetEq, version 4. It has been through extensive internal review during the course of the project. TEST=trybots Review URL: https://webrtc-codereview.appspot.com/1073005 git-svn-id: http://webrtc.googlecode.com/svn/trunk@3425 4adac7df-926f-26a2-2b94-8c16560cd09d
353 lines
12 KiB
C++
353 lines
12 KiB
C++
/*
|
|
* Copyright (c) 2012 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_coding/neteq4/dsp_helper.h"
|
|
|
|
#include <assert.h>
|
|
|
|
#include <algorithm> // Access to min, max.
|
|
#include <cstring> // Access to memset.
|
|
|
|
#include "webrtc/common_audio/signal_processing/include/signal_processing_library.h"
|
|
|
|
namespace webrtc {
|
|
|
|
// Table of constants used in method DspHelper::ParabolicFit().
|
|
const int16_t DspHelper::kParabolaCoefficients[17][3] = {
|
|
{ 120, 32, 64 },
|
|
{ 140, 44, 75 },
|
|
{ 150, 50, 80 },
|
|
{ 160, 57, 85 },
|
|
{ 180, 72, 96 },
|
|
{ 200, 89, 107 },
|
|
{ 210, 98, 112 },
|
|
{ 220, 108, 117 },
|
|
{ 240, 128, 128 },
|
|
{ 260, 150, 139 },
|
|
{ 270, 162, 144 },
|
|
{ 280, 174, 149 },
|
|
{ 300, 200, 160 },
|
|
{ 320, 228, 171 },
|
|
{ 330, 242, 176 },
|
|
{ 340, 257, 181 },
|
|
{ 360, 288, 192 } };
|
|
|
|
// Filter coefficients used when downsampling from the indicated sample rates
|
|
// (8, 16, 32, 48 kHz) to 4 kHz. Coefficients are in Q12. The corresponding Q0
|
|
// values are provided in the comments before each array.
|
|
|
|
// Q0 values: {0.3, 0.4, 0.3}.
|
|
const int16_t DspHelper::kDownsample8kHzTbl[3] = { 1229, 1638, 1229 };
|
|
|
|
// Q0 values: {0.15, 0.2, 0.3, 0.2, 0.15}.
|
|
const int16_t DspHelper::kDownsample16kHzTbl[5] = { 614, 819, 1229, 819, 614 };
|
|
|
|
// Q0 values: {0.1425, 0.1251, 0.1525, 0.1628, 0.1525, 0.1251, 0.1425}.
|
|
const int16_t DspHelper::kDownsample32kHzTbl[7] = {
|
|
584, 512, 625, 667, 625, 512, 584 };
|
|
|
|
// Q0 values: {0.2487, 0.0952, 0.1042, 0.1074, 0.1042, 0.0952, 0.2487}.
|
|
const int16_t DspHelper::kDownsample48kHzTbl[7] = {
|
|
1019, 390, 427, 440, 427, 390, 1019 };
|
|
|
|
int DspHelper::RampSignal(const int16_t* input,
|
|
size_t length,
|
|
int factor,
|
|
int increment,
|
|
int16_t* output) {
|
|
int factor_q20 = (factor << 6) + 32;
|
|
// TODO(hlundin): Add 32 to factor_q20 when converting back to Q14?
|
|
for (size_t i = 0; i < length; ++i) {
|
|
output[i] = (factor * input[i] + 8192) >> 14;
|
|
factor_q20 += increment;
|
|
factor_q20 = std::max(factor_q20, 0); // Never go negative.
|
|
factor = std::min(factor_q20 >> 6, 16384);
|
|
}
|
|
return factor;
|
|
}
|
|
|
|
int DspHelper::RampSignal(int16_t* signal,
|
|
size_t length,
|
|
int factor,
|
|
int increment) {
|
|
return RampSignal(signal, length, factor, increment, signal);
|
|
}
|
|
|
|
int DspHelper::RampSignal(AudioMultiVector<int16_t>* signal,
|
|
size_t start_index,
|
|
size_t length,
|
|
int factor,
|
|
int increment) {
|
|
assert(start_index + length <= signal->Size());
|
|
if (start_index + length > signal->Size()) {
|
|
// Wrong parameters. Do nothing and return the scale factor unaltered.
|
|
return factor;
|
|
}
|
|
int end_factor = 0;
|
|
// Loop over the channels, starting at the same |factor| each time.
|
|
for (size_t channel = 0; channel < signal->Channels(); ++channel) {
|
|
end_factor =
|
|
RampSignal(&(*signal)[channel][start_index], length, factor, increment);
|
|
}
|
|
return end_factor;
|
|
}
|
|
|
|
void DspHelper::PeakDetection(int16_t* data, int data_length,
|
|
int num_peaks, int fs_mult,
|
|
int* peak_index, int16_t* peak_value) {
|
|
int16_t min_index = 0;
|
|
int16_t max_index = 0;
|
|
|
|
for (int i = 0; i <= num_peaks - 1; i++) {
|
|
if (num_peaks == 1) {
|
|
// Single peak. The parabola fit assumes that an extra point is
|
|
// available; worst case it gets a zero on the high end of the signal.
|
|
// TODO(hlundin): This can potentially get much worse. It breaks the
|
|
// API contract, that the length of |data| is |data_length|.
|
|
data_length++;
|
|
}
|
|
|
|
peak_index[i] = WebRtcSpl_MaxIndexW16(data, data_length - 1);
|
|
|
|
if (i != num_peaks - 1) {
|
|
min_index = std::max(0, peak_index[i] - 2);
|
|
max_index = std::min(data_length - 1, peak_index[i] + 2);
|
|
}
|
|
|
|
if ((peak_index[i] != 0) && (peak_index[i] != (data_length - 2))) {
|
|
ParabolicFit(&data[peak_index[i] - 1], fs_mult, &peak_index[i],
|
|
&peak_value[i]);
|
|
} else {
|
|
if (peak_index[i] == data_length - 2) {
|
|
if (data[peak_index[i]] > data[peak_index[i] + 1]) {
|
|
ParabolicFit(&data[peak_index[i] - 1], fs_mult, &peak_index[i],
|
|
&peak_value[i]);
|
|
} else if (data[peak_index[i]] <= data[peak_index[i] + 1]) {
|
|
// Linear approximation.
|
|
peak_value[i] = (data[peak_index[i]] + data[peak_index[i] + 1]) >> 1;
|
|
peak_index[i] = (peak_index[i] * 2 + 1) * fs_mult;
|
|
}
|
|
} else {
|
|
peak_value[i] = data[peak_index[i]];
|
|
peak_index[i] = peak_index[i] * 2 * fs_mult;
|
|
}
|
|
}
|
|
|
|
if (i != num_peaks - 1) {
|
|
memset(&data[min_index], 0,
|
|
sizeof(data[0]) * (max_index - min_index + 1));
|
|
}
|
|
}
|
|
}
|
|
|
|
void DspHelper::ParabolicFit(int16_t* signal_points, int fs_mult,
|
|
int* peak_index, int16_t* peak_value) {
|
|
uint16_t fit_index[13];
|
|
if (fs_mult == 1) {
|
|
fit_index[0] = 0;
|
|
fit_index[1] = 8;
|
|
fit_index[2] = 16;
|
|
} else if (fs_mult == 2) {
|
|
fit_index[0] = 0;
|
|
fit_index[1] = 4;
|
|
fit_index[2] = 8;
|
|
fit_index[3] = 12;
|
|
fit_index[4] = 16;
|
|
} else if (fs_mult == 4) {
|
|
fit_index[0] = 0;
|
|
fit_index[1] = 2;
|
|
fit_index[2] = 4;
|
|
fit_index[3] = 6;
|
|
fit_index[4] = 8;
|
|
fit_index[5] = 10;
|
|
fit_index[6] = 12;
|
|
fit_index[7] = 14;
|
|
fit_index[8] = 16;
|
|
} else {
|
|
fit_index[0] = 0;
|
|
fit_index[1] = 1;
|
|
fit_index[2] = 3;
|
|
fit_index[3] = 4;
|
|
fit_index[4] = 5;
|
|
fit_index[5] = 7;
|
|
fit_index[6] = 8;
|
|
fit_index[7] = 9;
|
|
fit_index[8] = 11;
|
|
fit_index[9] = 12;
|
|
fit_index[10] = 13;
|
|
fit_index[11] = 15;
|
|
fit_index[12] = 16;
|
|
}
|
|
|
|
// num = -3 * signal_points[0] + 4 * signal_points[1] - signal_points[2];
|
|
// den = signal_points[0] - 2 * signal_points[1] + signal_points[2];
|
|
int32_t num = (signal_points[0] * -3) + (signal_points[1] * 4)
|
|
- signal_points[2];
|
|
int32_t den = signal_points[0] + (signal_points[1] * -2) + signal_points[2];
|
|
int32_t temp = num * 120;
|
|
int flag = 1;
|
|
int16_t stp = kParabolaCoefficients[fit_index[fs_mult]][0]
|
|
- kParabolaCoefficients[fit_index[fs_mult - 1]][0];
|
|
int16_t strt = (kParabolaCoefficients[fit_index[fs_mult]][0]
|
|
+ kParabolaCoefficients[fit_index[fs_mult - 1]][0]) / 2;
|
|
int16_t lmt;
|
|
if (temp < -den * strt) {
|
|
lmt = strt - stp;
|
|
while (flag) {
|
|
if ((flag == fs_mult) || (temp > -den * lmt)) {
|
|
*peak_value = (den * kParabolaCoefficients[fit_index[fs_mult - flag]][1]
|
|
+ num * kParabolaCoefficients[fit_index[fs_mult - flag]][2]
|
|
+ signal_points[0] * 256) / 256;
|
|
*peak_index = *peak_index * 2 * fs_mult - flag;
|
|
flag = 0;
|
|
} else {
|
|
flag++;
|
|
lmt -= stp;
|
|
}
|
|
}
|
|
} else if (temp > -den * (strt + stp)) {
|
|
lmt = strt + 2 * stp;
|
|
while (flag) {
|
|
if ((flag == fs_mult) || (temp < -den * lmt)) {
|
|
int32_t temp_term_1 =
|
|
den * kParabolaCoefficients[fit_index[fs_mult+flag]][1];
|
|
int32_t temp_term_2 =
|
|
num * kParabolaCoefficients[fit_index[fs_mult+flag]][2];
|
|
int32_t temp_term_3 = signal_points[0] * 256;
|
|
*peak_value = (temp_term_1 + temp_term_2 + temp_term_3) / 256;
|
|
*peak_index = *peak_index * 2 * fs_mult + flag;
|
|
flag = 0;
|
|
} else {
|
|
flag++;
|
|
lmt += stp;
|
|
}
|
|
}
|
|
} else {
|
|
*peak_value = signal_points[1];
|
|
*peak_index = *peak_index * 2 * fs_mult;
|
|
}
|
|
}
|
|
|
|
int DspHelper::MinDistortion(const int16_t* signal, int min_lag,
|
|
int max_lag, int length,
|
|
int32_t* distortion_value) {
|
|
int best_index = -1;
|
|
int32_t min_distortion = WEBRTC_SPL_WORD32_MAX;
|
|
for (int i = min_lag; i <= max_lag; i++) {
|
|
int32_t sum_diff = 0;
|
|
const int16_t* data1 = signal;
|
|
const int16_t* data2 = signal - i;
|
|
for (int j = 0; j < length; j++) {
|
|
sum_diff += WEBRTC_SPL_ABS_W32(data1[j] - data2[j]);
|
|
}
|
|
// Compare with previous minimum.
|
|
if (sum_diff < min_distortion) {
|
|
min_distortion = sum_diff;
|
|
best_index = i;
|
|
}
|
|
}
|
|
*distortion_value = min_distortion;
|
|
return best_index;
|
|
}
|
|
|
|
void DspHelper::CrossFade(const int16_t* input1, const int16_t* input2,
|
|
int length, int16_t* mix_factor,
|
|
int16_t factor_decrement, int16_t* output) {
|
|
int16_t factor = *mix_factor;
|
|
int16_t complement_factor = 16384 - factor;
|
|
for (int i = 0; i < length; i++) {
|
|
output[i] =
|
|
(factor * input1[i] + complement_factor * input2[i] + 8192) >> 14;
|
|
factor -= factor_decrement;
|
|
complement_factor += factor_decrement;
|
|
}
|
|
*mix_factor = factor;
|
|
}
|
|
|
|
void DspHelper::UnmuteSignal(const int16_t* input, int length, int16_t* factor,
|
|
int16_t increment, int16_t* output) {
|
|
uint16_t factor_16b = *factor;
|
|
int32_t factor_32b = (static_cast<int32_t>(factor_16b) << 6) + 32;
|
|
for (int i = 0; i < length; i++) {
|
|
output[i] = (factor_16b * input[i] + 8192) >> 14;
|
|
factor_32b = std::max(factor_32b + increment, 0);
|
|
factor_16b = std::min(16384, factor_32b >> 6);
|
|
}
|
|
*factor = factor_16b;
|
|
}
|
|
|
|
void DspHelper::MuteSignal(int16_t* signal, int16_t mute_slope, int length) {
|
|
int32_t factor = (16384 << 6) + 32;
|
|
for (int i = 0; i < length; i++) {
|
|
signal[i] = ((factor >> 6) * signal[i] + 8192) >> 14;
|
|
factor -= mute_slope;
|
|
}
|
|
}
|
|
|
|
int DspHelper::DownsampleTo4kHz(const int16_t* input, int input_length,
|
|
int output_length, int input_rate_hz,
|
|
bool compensate_delay, int16_t* output) {
|
|
// Set filter parameters depending on input frequency.
|
|
// NOTE: The phase delay values are wrong compared to the true phase delay
|
|
// of the filters. However, the error is preserved (through the +1 term) for
|
|
// consistency.
|
|
const int16_t* filter_coefficients; // Filter coefficients.
|
|
int16_t filter_length; // Number of coefficients.
|
|
int16_t filter_delay; // Phase delay in samples.
|
|
int16_t factor; // Conversion rate (inFsHz / 8000).
|
|
switch (input_rate_hz) {
|
|
case 8000: {
|
|
filter_length = 3;
|
|
factor = 2;
|
|
filter_coefficients = kDownsample8kHzTbl;
|
|
filter_delay = 1 + 1;
|
|
break;
|
|
}
|
|
case 16000: {
|
|
filter_length = 5;
|
|
factor = 4;
|
|
filter_coefficients = kDownsample16kHzTbl;
|
|
filter_delay = 2 + 1;
|
|
break;
|
|
}
|
|
case 32000: {
|
|
filter_length = 7;
|
|
factor = 8;
|
|
filter_coefficients = kDownsample32kHzTbl;
|
|
filter_delay = 3 + 1;
|
|
break;
|
|
}
|
|
case 48000: {
|
|
filter_length = 7;
|
|
factor = 12;
|
|
filter_coefficients = kDownsample48kHzTbl;
|
|
filter_delay = 3 + 1;
|
|
break;
|
|
}
|
|
default: {
|
|
assert(false);
|
|
return -1;
|
|
}
|
|
}
|
|
|
|
if (!compensate_delay) {
|
|
// Disregard delay compensation.
|
|
filter_delay = 0;
|
|
}
|
|
|
|
// Returns -1 if input signal is too short; 0 otherwise.
|
|
return WebRtcSpl_DownsampleFast(&input[filter_length - 1],
|
|
input_length - (filter_length - 1), output,
|
|
output_length, filter_coefficients,
|
|
filter_length, factor, filter_delay);
|
|
}
|
|
|
|
} // namespace webrtc
|