Bump to WebRTC M120 release

Some API deprecation -- ExperimentalAgc and ExperimentalNs are gone.
We're continuing to carry iSAC even though it's gone upstream, but maybe
we'll want to drop that soon.
This commit is contained in:
Arun Raghavan
2023-12-12 10:42:58 -05:00
parent 9a202fb8c2
commit c6abf6cd3f
479 changed files with 20900 additions and 11996 deletions

View File

@ -16,27 +16,23 @@
#include <numeric>
#include "rtc_base/checks.h"
#include "rtc_base/numerics/safe_compare.h"
namespace webrtc {
namespace rnn_vad {
namespace {
// Computes cross-correlation coefficients between |x| and |y| and writes them
// in |x_corr|. The lag values are in {0, ..., max_lag - 1}, where max_lag
// equals the size of |x_corr|.
// The |x| and |y| sub-arrays used to compute a cross-correlation coefficients
// for a lag l have both size "size of |x| - l" - i.e., the longest sub-array is
// used. |x| and |y| must have the same size.
void ComputeCrossCorrelation(
// Computes auto-correlation coefficients for `x` and writes them in
// `auto_corr`. The lag values are in {0, ..., max_lag - 1}, where max_lag
// equals the size of `auto_corr`.
void ComputeAutoCorrelation(
rtc::ArrayView<const float> x,
rtc::ArrayView<const float> y,
rtc::ArrayView<float, kNumLpcCoefficients> x_corr) {
constexpr size_t max_lag = x_corr.size();
RTC_DCHECK_EQ(x.size(), y.size());
rtc::ArrayView<float, kNumLpcCoefficients> auto_corr) {
constexpr int max_lag = auto_corr.size();
RTC_DCHECK_LT(max_lag, x.size());
for (size_t lag = 0; lag < max_lag; ++lag) {
x_corr[lag] =
std::inner_product(x.begin(), x.end() - lag, y.begin() + lag, 0.f);
for (int lag = 0; lag < max_lag; ++lag) {
auto_corr[lag] =
std::inner_product(x.begin(), x.end() - lag, x.begin() + lag, 0.f);
}
}
@ -45,9 +41,13 @@ void DenoiseAutoCorrelation(
rtc::ArrayView<float, kNumLpcCoefficients> auto_corr) {
// Assume -40 dB white noise floor.
auto_corr[0] *= 1.0001f;
for (size_t i = 1; i < kNumLpcCoefficients; ++i) {
auto_corr[i] -= auto_corr[i] * (0.008f * i) * (0.008f * i);
}
// Hard-coded values obtained as
// [np.float32((0.008*0.008*i*i)) for i in range(1,5)].
auto_corr[1] -= auto_corr[1] * 0.000064f;
auto_corr[2] -= auto_corr[2] * 0.000256f;
auto_corr[3] -= auto_corr[3] * 0.000576f;
auto_corr[4] -= auto_corr[4] * 0.001024f;
static_assert(kNumLpcCoefficients == 5, "Update `auto_corr`.");
}
// Computes the initial inverse filter coefficients given the auto-correlation
@ -56,9 +56,9 @@ void ComputeInitialInverseFilterCoefficients(
rtc::ArrayView<const float, kNumLpcCoefficients> auto_corr,
rtc::ArrayView<float, kNumLpcCoefficients - 1> lpc_coeffs) {
float error = auto_corr[0];
for (size_t i = 0; i < kNumLpcCoefficients - 1; ++i) {
for (int i = 0; i < kNumLpcCoefficients - 1; ++i) {
float reflection_coeff = 0.f;
for (size_t j = 0; j < i; ++j) {
for (int j = 0; j < i; ++j) {
reflection_coeff += lpc_coeffs[j] * auto_corr[i - j];
}
reflection_coeff += auto_corr[i + 1];
@ -72,7 +72,7 @@ void ComputeInitialInverseFilterCoefficients(
reflection_coeff /= -error;
// Update LPC coefficients and total error.
lpc_coeffs[i] = reflection_coeff;
for (size_t j = 0; j<(i + 1)>> 1; ++j) {
for (int j = 0; j < ((i + 1) >> 1); ++j) {
const float tmp1 = lpc_coeffs[j];
const float tmp2 = lpc_coeffs[i - 1 - j];
lpc_coeffs[j] = tmp1 + reflection_coeff * tmp2;
@ -91,46 +91,49 @@ void ComputeAndPostProcessLpcCoefficients(
rtc::ArrayView<const float> x,
rtc::ArrayView<float, kNumLpcCoefficients> lpc_coeffs) {
std::array<float, kNumLpcCoefficients> auto_corr;
ComputeCrossCorrelation(x, x, {auto_corr.data(), auto_corr.size()});
ComputeAutoCorrelation(x, auto_corr);
if (auto_corr[0] == 0.f) { // Empty frame.
std::fill(lpc_coeffs.begin(), lpc_coeffs.end(), 0);
return;
}
DenoiseAutoCorrelation({auto_corr.data(), auto_corr.size()});
DenoiseAutoCorrelation(auto_corr);
std::array<float, kNumLpcCoefficients - 1> lpc_coeffs_pre{};
ComputeInitialInverseFilterCoefficients(auto_corr, lpc_coeffs_pre);
// LPC coefficients post-processing.
// TODO(bugs.webrtc.org/9076): Consider removing these steps.
float c1 = 1.f;
for (size_t i = 0; i < kNumLpcCoefficients - 1; ++i) {
c1 *= 0.9f;
lpc_coeffs_pre[i] *= c1;
}
const float c2 = 0.8f;
lpc_coeffs[0] = lpc_coeffs_pre[0] + c2;
lpc_coeffs[1] = lpc_coeffs_pre[1] + c2 * lpc_coeffs_pre[0];
lpc_coeffs[2] = lpc_coeffs_pre[2] + c2 * lpc_coeffs_pre[1];
lpc_coeffs[3] = lpc_coeffs_pre[3] + c2 * lpc_coeffs_pre[2];
lpc_coeffs[4] = c2 * lpc_coeffs_pre[3];
lpc_coeffs_pre[0] *= 0.9f;
lpc_coeffs_pre[1] *= 0.9f * 0.9f;
lpc_coeffs_pre[2] *= 0.9f * 0.9f * 0.9f;
lpc_coeffs_pre[3] *= 0.9f * 0.9f * 0.9f * 0.9f;
constexpr float kC = 0.8f;
lpc_coeffs[0] = lpc_coeffs_pre[0] + kC;
lpc_coeffs[1] = lpc_coeffs_pre[1] + kC * lpc_coeffs_pre[0];
lpc_coeffs[2] = lpc_coeffs_pre[2] + kC * lpc_coeffs_pre[1];
lpc_coeffs[3] = lpc_coeffs_pre[3] + kC * lpc_coeffs_pre[2];
lpc_coeffs[4] = kC * lpc_coeffs_pre[3];
static_assert(kNumLpcCoefficients == 5, "Update `lpc_coeffs(_pre)`.");
}
void ComputeLpResidual(
rtc::ArrayView<const float, kNumLpcCoefficients> lpc_coeffs,
rtc::ArrayView<const float> x,
rtc::ArrayView<float> y) {
RTC_DCHECK_LT(kNumLpcCoefficients, x.size());
RTC_DCHECK_GT(x.size(), kNumLpcCoefficients);
RTC_DCHECK_EQ(x.size(), y.size());
std::array<float, kNumLpcCoefficients> input_chunk;
input_chunk.fill(0.f);
for (size_t i = 0; i < y.size(); ++i) {
const float sum = std::inner_product(input_chunk.begin(), input_chunk.end(),
lpc_coeffs.begin(), x[i]);
// Circular shift and add a new sample.
for (size_t j = kNumLpcCoefficients - 1; j > 0; --j)
input_chunk[j] = input_chunk[j - 1];
input_chunk[0] = x[i];
// Copy result.
y[i] = sum;
// The code below implements the following operation:
// y[i] = x[i] + dot_product({x[i], ..., x[i - kNumLpcCoefficients + 1]},
// lpc_coeffs)
// Edge case: i < kNumLpcCoefficients.
y[0] = x[0];
for (int i = 1; i < kNumLpcCoefficients; ++i) {
y[i] =
std::inner_product(x.crend() - i, x.crend(), lpc_coeffs.cbegin(), x[i]);
}
// Regular case.
auto last = x.crend();
for (int i = kNumLpcCoefficients; rtc::SafeLt(i, y.size()); ++i, --last) {
y[i] = std::inner_product(last - kNumLpcCoefficients, last,
lpc_coeffs.cbegin(), x[i]);
}
}