Skip to content

fix(windows): low-pass before decimating so 96/192 kHz sources stop aliasing #644

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
My-Denia wants to merge 2 commits into
base: main
Choose a base branch
from
+1,239 −78
Open

fix(windows): low-pass before decimating so 96/192 kHz sources stop aliasing #644

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
8d97f79
fix(windows): low-pass before decimating so 96/192 kHz sources stop a…
My-Denia Sep 10, 2026
0481090
docs(testing): add the anti-alias check, and log the A/B that confirm…
My-Denia Sep 10, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
236 changes: 185 additions & 51 deletions electron/native/wgc-capture/src/audio_sample_utils.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,6 +5,7 @@
#include <algorithm>
#include <chrono>
#include <cmath>
#include <numeric>
#include <cstddef>
#include <cstring>
#include <limits>
Expand Down Expand Up @@ -133,6 +134,120 @@ AudioInputFormat makeAacCompatibleAudioFormat(const AudioInputFormat& source) {
return format;
}

namespace {

// Zeroth-order modified Bessel of the first kind, for the Kaiser window. The
// series converges fast for the beta this file uses; the loop bails once a term
// stops moving the sum at double precision.
constexpr double kPi = 3.14159265358979323846;

double besselI0(double x) {
double sum = 1.0;
double term = 1.0;
for (int k = 1; k < 128; k += 1) {
term *= (x * x / 4.0) / (static_cast<double>(k) * k);
sum += term;
if (term < sum * 1e-17) {
break;
}
}
return sum;
}

// The whole anti-alias filter is these three numbers, and two properties follow
// from them for EVERY decimation factor -- which is what keeps this path general
// instead of tuned for the two rates that happen to be common:
//
// * the cutoff always lands at 11/12 = 0.9167 of the OUTPUT Nyquist, because
// both it and the Nyquist scale as 1/factor. So the passband stays flat to
// roughly 0.8 of the output Nyquist and the stop band is already deep by the
// time anything can fold back;
// * the group delay is always 80*factor/2 source frames. After the decimation
// phase hands back (factor-1) of them, the net is 39 + 1/factor output
// frames -- bounded in [39, 39.5] whatever the factor, about 0.82 ms at
// 48 kHz out.
//
// Depth: a 36 kHz tone leaves 96 kHz -> 48 kHz at about -117 dB, against roughly
// -8 dB for the box average this replaces. 80 taps per step is ~1.5x the Kaiser
// minimum for that transition and depth. The margin is affordable because the
// cost is set by the SOURCE rate rather than the factor: the tap count grows
// with factor exactly as fast as the output rate falls, so 192 kHz -> 48 kHz and
// 192 kHz -> 8 kHz cost the same.
constexpr size_t kTapsPerDecimationStep = 80;
constexpr double kCutoffFractionOfSourceRate = 11.0 / 24.0;
constexpr double kKaiserBeta = 8.6;

} // namespace

void AudioDecimatorState::reset() {
std::fill(history_.begin(), history_.end(), 0.0);
position_ = 0;
phase_ = 0;
}

void AudioDecimatorState::prepare(UINT32 factor, UINT32 channels) {
if (factor_ == factor && channels_ == channels && !taps_.empty()) {
return;
}
factor_ = factor;
channels_ = channels;
// Kaiser-windowed sinc. The three constants and what follows from them are
// documented where they are defined; here it is enough that the cutoff is
// expressed against the SOURCE rate, which is what makes every factor land
// on the same fraction of its own output Nyquist.
const size_t tapCount = kTapsPerDecimationStep * factor + 1;
const double cutoff = kCutoffFractionOfSourceRate / static_cast<double>(factor);
const double middle = static_cast<double>(tapCount - 1) / 2.0;
const double norm = besselI0(kKaiserBeta);
taps_.assign(tapCount, 0.0);
for (size_t k = 0; k < tapCount; k += 1) {
const double offset = static_cast<double>(k) - middle;
const double sinc = std::abs(offset) < 1e-12
? 2.0 * cutoff
: std::sin(2.0 * kPi * cutoff * offset) / (kPi * offset);
const double ratio = offset / middle;
taps_[k] =
sinc * besselI0(kKaiserBeta * std::sqrt(std::max(0.0, 1.0 - ratio * ratio))) / norm;
}
// Unity at DC, so the passband keeps the level the caller handed us.
const double dc = std::accumulate(taps_.begin(), taps_.end(), 0.0);
if (dc != 0.0) {
for (double& tap : taps_) {
tap /= dc;
}
}
history_.assign(tapCount * channels, 0.0);
position_ = 0;
phase_ = 0;
}

bool AudioDecimatorState::consume(const double* frame, double* out) {
if (taps_.empty() || channels_ == 0) {
return false;
}
for (UINT32 channel = 0; channel < channels_; channel += 1) {
history_[position_ * channels_ + channel] = frame[channel];
}
bool produced = false;
phase_ += 1;
if (phase_ == factor_) {
phase_ = 0;
produced = true;
const size_t tapCount = taps_.size();
for (UINT32 channel = 0; channel < channels_; channel += 1) {
double sum = 0.0;
size_t read = position_;
for (size_t k = 0; k < tapCount; k += 1) {
sum += taps_[k] * history_[read * channels_ + channel];
read = read != 0 ? read - 1 : tapCount - 1;
}
out[channel] = sum;
}
}
position_ = (position_ + 1 == taps_.size()) ? 0 : position_ + 1;
return produced;
}

void copyAudioWithGain(
const BYTE* source,
DWORD byteCount,
Expand Down Expand Up @@ -190,9 +305,9 @@ void convertAudioWithGain(
const AudioInputFormat& targetFormat,
double gain,
std::vector<BYTE>& destination) {
std::vector<BYTE> discardedRemainder;
AudioDecimatorState discardedDecimator;
convertAudioWithGain(
source, byteCount, sourceFormat, targetFormat, gain, destination, discardedRemainder);
source, byteCount, sourceFormat, targetFormat, gain, destination, discardedDecimator);
}

void convertAudioWithGain(
Expand All @@ -202,7 +317,7 @@ void convertAudioWithGain(
const AudioInputFormat& targetFormat,
double gain,
std::vector<BYTE>& destination,
std::vector<BYTE>& remainder) {
AudioDecimatorState& decimator) {
if (!source || byteCount == 0 || sourceFormat.blockAlign == 0 || targetFormat.blockAlign == 0 ||
sourceFormat.sampleRate == 0 || targetFormat.sampleRate == 0 || sourceFormat.channels == 0 ||
targetFormat.channels == 0) {
Expand All @@ -211,6 +326,20 @@ void convertAudioWithGain(
}

if (sameAudioFormatForMixing(sourceFormat, targetFormat)) {
// The decimator belongs to the caller, not to this call, so a packet
// that does not decimate has to hand it back empty rather than leave a
// half-finished group waiting for frames that will never arrive.
//
// In this helper that is a contract, not a live path. The mix format is
// read once per session (a single GetMixFormat in
// WasapiLoopbackCapture) and resolveInputFormat passes its subtype
// through unchanged, so what
// arrives here is whatever the audio engine hands out -- which is
// float32, while the encoder target is always PCM16. Subtype alone
// therefore never matches, and this branch is unreachable in
// production. The suite does reach it, by reusing one state across
// formats -- which is the point of writing it.
decimator.reset();
copyAudioWithGain(source, byteCount, targetFormat, gain, destination);
return;
}
Expand All @@ -221,58 +350,63 @@ void convertAudioWithGain(
return;
}

// Integer-factor downsample (96 kHz / 192 kHz -> 48 kHz): average each
// group of source frames instead of picking one. Nearest-neighbour
// decimation aliases content above the new Nyquist into the recording.
// Incomplete groups stay in remainder so the next packet can finish them.
// A note on level, because this filter can do something the box average could
// not: a decimated peak can exceed the largest input sample it came from. A
// transition band this sharp needs negative taps, and negative taps make the
// step response overshoot; the box average never overshoots only because its
// taps are all positive, which is also why it rejects so little. HOW MUCH it
// overshoots is this design's choice, set by the cutoff and the window. For
// a near-full-scale 1 kHz square wave -- the Gibbs case -- it is +1.4 to
// +1.5 dB depending on the factor; a square with a higher fundamental
// overshoots more, up to about +3.8 dB in the cases measured. For arbitrary
// bounded input the ceiling is the sum of |taps|, 2.22 to 2.23 depending on
// the factor, or about +7 dB. writeSampleFromDouble clamps, so overshoot
// saturates rather than wrapping; the suite checks that write sample by
// sample, and pins the 1 kHz figure at every factor it tests.

// Integer-factor downsample (96 kHz / 192 kHz -> 48 kHz). Dropping every
// Nth frame is only safe once the content above the new Nyquist is gone, so
// every frame runs through an anti-alias low-pass first and that filter's
// history rides across packets inside `decimator`. A group a packet leaves
// unfinished is no longer held back as bytes — its frames are already in the
// filter — and `pendingFrames()` is what reports how many there were.
if (sourceFormat.sampleRate > targetFormat.sampleRate &&
sourceFormat.sampleRate % targetFormat.sampleRate == 0) {
const UINT32 factor = sourceFormat.sampleRate / targetFormat.sampleRate;
if (remainder.size() % sourceFormat.blockAlign != 0) {
remainder.clear();
}
std::vector<BYTE> combined;
combined.reserve(remainder.size() + byteCount);
combined.insert(combined.end(), remainder.begin(), remainder.end());
combined.insert(combined.end(), source, source + byteCount);
const size_t totalFrames = combined.size() / sourceFormat.blockAlign;
const size_t targetFrames = totalFrames / factor;
const size_t consumedFrames = targetFrames * factor;
const size_t leftoverBytes = (totalFrames - consumedFrames) * sourceFormat.blockAlign;
if (targetFrames == 0) {
destination.clear();
remainder.swap(combined);
return;
}
destination.assign(targetFrames * targetFormat.blockAlign, 0);
for (size_t targetFrame = 0; targetFrame < targetFrames; ++targetFrame) {
decimator.prepare(factor, targetFormat.channels);
const size_t maxTargetFrames = (packetFrames + decimator.pendingFrames()) / factor;
destination.assign(maxTargetFrames * targetFormat.blockAlign, 0);
std::vector<double> frame(targetFormat.channels, 0.0);
std::vector<double> filtered(targetFormat.channels, 0.0);
size_t targetFrame = 0;
for (size_t sourceFrame = 0; sourceFrame < packetFrames; ++sourceFrame) {
for (UINT32 channel = 0; channel < targetFormat.channels; ++channel) {
frame[channel] = readMappedChannel(
source, sourceFormat, sourceFrame, channel, targetFormat.channels);
}
if (!decimator.consume(frame.data(), filtered.data())) {
continue;
}
for (UINT32 channel = 0; channel < targetFormat.channels; ++channel) {
double sum = 0.0;
for (UINT32 tap = 0; tap < factor; ++tap) {
sum += readMappedChannel(
combined.data(),
sourceFormat,
targetFrame * factor + tap,
channel,
targetFormat.channels);
}
writeSampleFromDouble(
destination.data(),
targetFormat,
targetFrame,
channel,
(sum / static_cast<double>(factor)) * gain);
filtered[channel] * gain);
}
targetFrame += 1;
}
remainder.assign(
combined.begin() + static_cast<std::ptrdiff_t>(consumedFrames * sourceFormat.blockAlign),
combined.end());
if (remainder.size() != leftoverBytes) {
remainder.resize(leftoverBytes);
}
destination.resize(targetFrame * targetFormat.blockAlign);
return;
}

// Same ownership rule as the pass-through above, but unlike that one this
// branch is hot. A 48 kHz microphone against a 48 kHz PCM16 target fails
// sameAudioFormatForMixing on subtype alone -- WASAPI hands out float32 --
// and `sourceRate > targetRate` is false, so on an ordinary machine every
// microphone packet lands here and resets a decimator it never used.
decimator.reset();
const size_t sourceFrames = packetFrames;
const double rateRatio = static_cast<double>(targetFormat.sampleRate) /
static_cast<double>(sourceFormat.sampleRate);
Expand Down Expand Up @@ -371,8 +505,8 @@ bool AudioMixer::start() {
emittedFrames_ = 0;
timelineStarted_ = false;
paused_ = false;
systemResampleRemainder_.clear();
microphoneResampleRemainder_.clear();
systemDecimator_.reset();
microphoneDecimator_.reset();
thread_ = std::thread([this] {
mixLoop();
});
Expand All @@ -384,8 +518,8 @@ void AudioMixer::beginTimeline() {
std::scoped_lock lock(mutex_);
systemQueue_.clear();
microphoneQueue_.clear();
systemResampleRemainder_.clear();
microphoneResampleRemainder_.clear();
systemDecimator_.reset();
microphoneDecimator_.reset();
emittedFrames_ = 0;
timelineStarted_ = true;
}
Expand All @@ -399,8 +533,8 @@ void AudioMixer::setPaused(bool paused) {
if (paused_) {
systemQueue_.clear();
microphoneQueue_.clear();
systemResampleRemainder_.clear();
microphoneResampleRemainder_.clear();
systemDecimator_.reset();
microphoneDecimator_.reset();
}
}
cv_.notify_all();
Expand All @@ -424,7 +558,7 @@ void AudioMixer::pushSystem(const BYTE* data, DWORD byteCount) {
if (paused_) {
return;
}
append(systemQueue_, data, byteCount, systemFormat_, 1.0, systemResampleRemainder_);
append(systemQueue_, data, byteCount, systemFormat_, 1.0, systemDecimator_);
}
cv_.notify_all();
}
Expand All @@ -445,7 +579,7 @@ void AudioMixer::pushMicrophone(const BYTE* data, DWORD byteCount) {
byteCount,
microphoneFormat_,
microphoneGain_,
microphoneResampleRemainder_);
microphoneDecimator_);
}
cv_.notify_all();
}
Expand All @@ -456,12 +590,12 @@ void AudioMixer::append(
DWORD byteCount,
const AudioInputFormat& sourceFormat,
double gain,
std::vector<BYTE>& remainder) {
AudioDecimatorState& decimator) {
if (!data || byteCount == 0) {
return;
}

convertAudioWithGain(data, byteCount, sourceFormat, format_, gain, gainBuffer_, remainder);
convertAudioWithGain(data, byteCount, sourceFormat, format_, gain, gainBuffer_, decimator);
queue.insert(queue.end(), gainBuffer_.begin(), gainBuffer_.end());
}

Expand Down
Loading
Loading