FEAT(client): Add interaural delay

There is a small time delay (interaural time delay or ITD) between your
ears depending on the sound source position on the horizontal plane and
the distance between your ears. This commit will add this delay by using
the extra sound data in the audio buffer.

For me at least, implementing this makes it much easier to identify
where a sound is coming from when positional audio is enabled.

Implements #2324

src/mumble/Audio.h

@@ -17,6 +17,12 @@
 
 #define SAMPLE_RATE 48000
 
+// interaural delay (in samples) for a sound coming directly from the side of the head
+// A Wikipedia article claims the average distance between ears is 15.2 cm for men
+// (0.44 ms) and 14.4 cm for women (0.42 ms). We decided to set the delay to 0.43 ms.
+// The delay is calculated from the distance and the speed of sound.
+constexpr float INTERAURAL_DELAY = 0.00043 / (1 / static_cast< float >(SAMPLE_RATE));
+
 typedef QPair< QString, QVariant > audioDevice;
 
 class LoopUser : public ClientUser {

src/mumble/AudioOutput.cpp

@@ -588,29 +588,59 @@ bool AudioOutput::mix(void *outbuff, unsigned int frameCount) {
 						aop->pfVolume[s] = -1.0;
 				}
 
+				if (!aop->piOffset) {
+					aop->piOffset = std::make_unique< unsigned int[] >(nchan);
+					for (unsigned int s = 0; s < nchan; ++s) {
+						aop->piOffset[s] = 0;
+					}
+				}
+
 				for (unsigned int s = 0; s < nchan; ++s) {
 					const float dot = bSpeakerPositional[s]
 										  ? connectionVec.x * speaker[s * 3 + 0] + connectionVec.y * speaker[s * 3 + 1]
 												+ connectionVec.z * speaker[s * 3 + 2]
 										  : 1.0f;
-					const float str   = svol[s] * calcGain(dot, len) * volumeAdjustment;
+					// Volume on the ear opposite to the sound should never reach 0 in the real world.
+					// The gain is multiplied by 19/20 and 1/20 is added. This will have the effect
+					// of bringing the lowest value up to 1/20, while keeping the highest value at 1.
+					// E.g. calcGain() = 1; 1 * 19/20 + 1/20 = 0.95 + 0.05 = 1
+					// calcGain() = 0; 0 * 19/20 + 1/20 = 0 + 0.05 = 0.05
+					const float str   = svol[s] * (1 / 20.0 + (19 / 20.0) * calcGain(dot, len)) * volumeAdjustment;
 					float *RESTRICT o = output + s;
 					const float old   = (aop->pfVolume[s] >= 0.0f) ? aop->pfVolume[s] : str;
 					const float inc   = (str - old) / static_cast< float >(frameCount);
 					aop->pfVolume[s]  = str;
+
+					// Calculates the ITD offset of the audio data this frame.
+					// Interaural Time Delay (ITD) is a small time delay between your ears
+					// depending on the sound source position on the horizonal plane and the
+					// distance between your ears.
+					//
+					// Offset for ITD is not applied directly, but rather the offset is interpolated
+					// linearly across the entire chunk, between the offset of the last chunk and the
+					// newly calculated offset for this chunk. This prevents clicking / buzzing when the
+					// audio source or camera is moving, because abruptly changing offsets (and thus
+					// abruptly changing the playback position) will create a clicking noise.
+					const int offset =
+						INTERAURAL_DELAY * (1.0 + dot) / 2.0; // Normalize dot to range [0,1] instead [-1,1]
+					const int oldOffset   = aop->piOffset[s];
+					const float incOffset = (offset - oldOffset) / static_cast< float >(frameCount);
+					aop->piOffset[s]      = offset;
 					/*
 										qWarning("%d: Pos %f %f %f : Dot %f Len %f Str %f", s, speaker[s*3+0],
 					   speaker[s*3+1], speaker[s*3+2], dot, len, str);
 					*/
 					if ((old >= 0.00000001f) || (str >= 0.00000001f)) {
 						for (unsigned int i = 0; i < frameCount; ++i) {
+							unsigned int currentOffset = oldOffset + incOffset * i;
 							if (speech && speech->bStereo) {
 								// Mix stereo user's stream into mono
 								// frame: for a stereo stream, the [LR] pair inside ...[LR]LRLRLR.... is a frame
-								o[i * nchan] += (pfBuffer[2 * i] / 2.0 + pfBuffer[2 * i + 1] / 2.0)
-												* (old + inc * static_cast< float >(i));
+								o[i * nchan] +=
+									(pfBuffer[2 * i + currentOffset] / 2.0 + pfBuffer[2 * i + currentOffset + 1] / 2.0)
+									* (old + inc * static_cast< float >(i));
 							} else {
-								o[i * nchan] += pfBuffer[i] * (old + inc * static_cast< float >(i));
+								o[i * nchan] += pfBuffer[i + currentOffset] * (old + inc * static_cast< float >(i));
 							}
 						}
 					}

src/mumble/AudioOutputSample.cpp

@@ -228,7 +228,8 @@ bool AudioOutputSample::prepareSampleBuffer(unsigned int frameCount) {
 	iLastConsume = sampleCount;
 
 	// Check if we can satisfy request with current buffer
-	if (iBufferFilled >= sampleCount)
+	// Maximum interaural delay is accounted for to prevent audio glitches
+	if (iBufferFilled >= sampleCount + INTERAURAL_DELAY)
 		return true;
 
 	// Calculate the required buffersize to hold the results
@@ -241,7 +242,7 @@ bool AudioOutputSample::prepareSampleBuffer(unsigned int frameCount) {
 	bool eof = false;
 	sf_count_t read;
 	do {
-		resizeBuffer(iBufferFilled + sampleCount);
+		resizeBuffer(iBufferFilled + sampleCount + INTERAURAL_DELAY);
 
 		// If we need to resample, write to the buffer on stack
 		float *pOut = (srs) ? fOut : pfBuffer + iBufferFilled;
@@ -270,7 +271,7 @@ bool AudioOutputSample::prepareSampleBuffer(unsigned int frameCount) {
 		}
 
 		iBufferFilled += outlen * channels;
-	} while (iBufferFilled < sampleCount);
+	} while (iBufferFilled < sampleCount + INTERAURAL_DELAY);
 
 	if (eof && !bEof) {
 		emit playbackFinished();

src/mumble/AudioOutputSpeech.cpp

@@ -231,15 +231,16 @@ bool AudioOutputSpeech::prepareSampleBuffer(unsigned int frameCount) {
 
 	iLastConsume = sampleCount;
 
-	if (iBufferFilled >= sampleCount)
+	// Maximum interaural delay is accounted for to prevent audio glitches
+	if (iBufferFilled >= sampleCount + INTERAURAL_DELAY)
 		return bLastAlive;
 
 	float *pOut;
 	bool nextalive = bLastAlive;
 
-	while (iBufferFilled < sampleCount) {
+	while (iBufferFilled < sampleCount + INTERAURAL_DELAY) {
 		int decodedSamples = iFrameSize;
-		resizeBuffer(iBufferFilled + iOutputSize);
+		resizeBuffer(iBufferFilled + iOutputSize + INTERAURAL_DELAY);
 		// TODO: allocating memory in the audio callback will crash mumble in some cases.
 		//       we need to initialize the buffer with an appropriate size when initializing
 		//       this class. See #4250.

src/mumble/AudioOutputUser.h

@@ -7,6 +7,7 @@
 #define MUMBLE_MUMBLE_AUDIOOUTPUTUSER_H_
 
 #include <QtCore/QObject>
+#include <memory>
 
 class AudioOutputUser : public QObject {
 private:
@@ -28,7 +29,8 @@ class AudioOutputUser : public QObject {
 	const QString qsName;
 	float *pfBuffer = nullptr;
 	float *pfVolume = nullptr;
-	float fPos[3]   = { 0.0, 0.0, 0.0 };
+	std::unique_ptr< unsigned int[] > piOffset;
+	float fPos[3] = { 0.0, 0.0, 0.0 };
 	bool bStereo;
 	virtual bool prepareSampleBuffer(unsigned int snum) = 0;
 };