AudioInput: ensure our encoders are reset correctly when using > 1 frame per packet.

When AudioInput is configured to use more than a single frame per
packet, the "!bPreviousVoice"-check in encodeOpusFrame() and
encodeCELTFrame() is not sufficient in determining when to reset
the encoder.

This is because bPreviousVoice refers to the previous *frame*.

But when AudioInput is configured to use multiple frames per
packet, we won't enter encodeOpusFrame() at the initial frame
where bPreviousVoice would be false.

The result is that we *never* reset the current encoder
when using more than one frame per packet.

This change adds a flag, bResetEncoder, to the AudioInput class.

Now, When AudioInput encounters a frame that is speech, and the
previous frame was not (bIsSpeech && !bPreviousVoice), it will
set the bResetEncoder flag.

The encoder is now reset correctly when a new voice stream begins,
just like it is for the single frame-per-packet case.

[Backported from 23fa9b3979c244f1bcdbc80972a88ef7144f8f4a]

2 files changed, 19 insertions, 2 deletions

diff --git a/src/mumble/AudioInput.cpp b/src/mumble/AudioInput.cpp
index eb6bb59be..9fb2a068b 100644
--- a/src/mumble/AudioInput.cpp
+++ b/src/mumble/AudioInput.cpp
@@ -144,6 +144,8 @@ AudioInput::AudioInput() : opusBuffer(g.s.iFramesPerPacket * (SAMPLE_RATE / 100)
 
 	bPreviousVoice = false;
 
+	bResetEncoder = true;
+
 	pfMicInput = pfEchoInput = pfOutput = NULL;
 
 	iBitrate = 0;
@@ -594,6 +596,8 @@ void AudioInput::resetAudioProcessor() {
 		sesEcho = NULL;
 	}
 
+	bResetEncoder = true;
+
 	bResetProcessor = false;
 }
 
@@ -675,8 +679,10 @@ bool AudioInput::selectCodec() {
 int AudioInput::encodeOpusFrame(short *source, int size, unsigned char *buffer) {
 	int len = 0;
 #ifdef USE_OPUS
-	if (!bPreviousVoice)
+	if (bResetEncoder) {
 		opus_encoder_ctl(opusState, OPUS_RESET_STATE, NULL);
+		bResetEncoder = false;
+	}
 
 	opus_encoder_ctl(opusState, OPUS_SET_BITRATE(iAudioQuality));
 
@@ -692,8 +698,10 @@ int AudioInput::encodeCELTFrame(short *psSource, unsigned char *buffer) {
 	if (!cCodec)
 		return len;
 
-	if (!bPreviousVoice)
+	if (bResetEncoder) {
 		cCodec->celt_encoder_ctl(ceEncoder, CELT_RESET_STATE);
+		bResetEncoder = false;
+	}
 
 	cCodec->celt_encoder_ctl(ceEncoder, CELT_SET_PREDICTION(0));
 
@@ -841,6 +849,10 @@ void AudioInput::encodeAudioFrame() {
 		speex_preprocess_ctl(sppPreprocess, SPEEX_PREPROCESS_SET_AGC_INCREMENT, &increment);
 	}
 
+	if (bIsSpeech && !bPreviousVoice) {
+		bResetEncoder = true;
+	}
+
 	tIdle.restart();
 
 	unsigned char buffer[512];
diff --git a/src/mumble/AudioInput.h b/src/mumble/AudioInput.h
index b6cce4169..80cb60374 100644
--- a/src/mumble/AudioInput.h
+++ b/src/mumble/AudioInput.h
@@ -122,6 +122,11 @@ class AudioInput : public QThread {
 		CELTCodec *cCodec;
 		CELTEncoder *ceEncoder;
 
+		/// bResetEncoder is a flag that notifies
+		/// our encoder functions that the encoder
+		/// needs to be reset.
+		bool bResetEncoder;
+
 		int iAudioQuality;
 		int iAudioFrames;