diff --git a/dom/media/AudioPacketizer.h b/dom/media/AudioPacketizer.h
index 18adedf24248..3a202f78b553 100644
--- a/dom/media/AudioPacketizer.h
+++ b/dom/media/AudioPacketizer.h
@@ -24,10 +24,13 @@ namespace mozilla {
  * Input and output, as well as length units in the public interface are
  * interleaved frames.
  *
- * Allocations of output buffer are performed by this class.  Buffers can simply
- * be delete-d.  This is because packets are intended to be sent off to
+ * Allocations of output buffer can be performed by this class.  Buffers can
+ * simply be delete-d.  This is because packets are intended to be sent off to
  * non-gecko code using normal pointers/length pairs
  *
+ * Alternatively, consumers can pass in a buffer in which the output is copied.
+ * The buffer needs to be large enough to store a packet worth of audio.
+ *
  * The implementation uses a circular buffer using absolute virtual indices.
  */
 template <typename InputType, typename OutputType>
@@ -98,6 +101,15 @@ public:
     uint32_t samplesNeeded = mPacketSize * mChannels;
     OutputType* out = new OutputType[samplesNeeded];
 
+    Output(out);
+
+    return out;
+  }
+
+  void Output(OutputType* aOutputBuffer)
+  {
+    uint32_t samplesNeeded = mPacketSize * mChannels;
+
     // Under-run. Pad the end of the buffer with silence.
     if (AvailableSamples() < samplesNeeded) {
 #ifdef LOG_PACKETIZER_UNDERRUN
@@ -108,26 +120,24 @@ public:
       NS_WARNING(buf);
 #endif
       uint32_t zeros = samplesNeeded - AvailableSamples();
-      PodZero(out + AvailableSamples(), zeros);
+      PodZero(aOutputBuffer + AvailableSamples(), zeros);
       samplesNeeded -= zeros;
     }
     if (ReadIndex() + samplesNeeded <= mLength) {
       ConvertAudioSamples<InputType,OutputType>(mStorage.get() + ReadIndex(),
-                                                out,
+                                                aOutputBuffer,
                                                 samplesNeeded);
     } else {
       uint32_t firstPartLength = mLength - ReadIndex();
       uint32_t secondPartLength = samplesNeeded - firstPartLength;
       ConvertAudioSamples<InputType, OutputType>(mStorage.get() + ReadIndex(),
-                                                 out,
+                                                 aOutputBuffer,
                                                  firstPartLength);
       ConvertAudioSamples<InputType, OutputType>(mStorage.get(),
-                                                 out + firstPartLength,
+                                                 aOutputBuffer + firstPartLength,
                                                  secondPartLength);
     }
     mReadIndex += samplesNeeded;
-
-    return out;
   }
 
   uint32_t PacketsAvailable() const {
diff --git a/media/webrtc/signaling/src/mediapipeline/MediaPipeline.cpp b/media/webrtc/signaling/src/mediapipeline/MediaPipeline.cpp
index da5fed010fd0..14bcb2e976fe 100644
--- a/media/webrtc/signaling/src/mediapipeline/MediaPipeline.cpp
+++ b/media/webrtc/signaling/src/mediapipeline/MediaPipeline.cpp
@@ -953,6 +953,8 @@ void MediaPipelineTransmit::PipelineListener::ProcessAudioChunk(
   // channels (since the WebRTC.org code below makes the assumption that the
   // input audio is either mono or stereo).
   uint32_t outputChannels = chunk.ChannelCount() == 1 ? 1 : 2;
+  // XXX[padenot] We could remove this allocation for the common case, or stick
+  // it on the object so it only happens once.
   nsAutoArrayPtr<int16_t> convertedSamples(
       new int16_t[chunk.mDuration * outputChannels]);
 
@@ -998,8 +1000,16 @@ void MediaPipelineTransmit::PipelineListener::ProcessAudioChunk(
   while (packetizer_->PacketsAvailable()) {
     uint32_t samplesPerPacket = packetizer_->PacketSize() *
                                 packetizer_->Channels();
-    conduit->SendAudioFrame(packetizer_->Output(),
-                            samplesPerPacket ,
+
+    // We know that webrtc.org's code going to copy the samples down the line,
+    // so we can just use a stack buffer here instead of malloc-ing.
+    // Max size given stereo is 480*2*2 = 1920 (48KHz)
+    const size_t AUDIO_SAMPLE_BUFFER_MAX = 1920;
+    int16_t packet[AUDIO_SAMPLE_BUFFER_MAX];
+
+    packetizer_->Output(packet);
+    conduit->SendAudioFrame(packet,
+                            samplesPerPacket,
                             rate, 0);
   }
 }
@@ -1329,7 +1339,6 @@ NotifyPull(MediaStreamGraph* graph, StreamTime desired_time) {
   // This comparison is done in total time to avoid accumulated roundoff errors.
   while (source_->TicksToTimeRoundDown(track_rate_, played_ticks_) <
          desired_time) {
-    // TODO(ekr@rtfm.com): Is there a way to avoid mallocating here?  Or reduce the size?
     // Max size given stereo is 480*2*2 = 1920 (48KHz)
     const size_t AUDIO_SAMPLE_BUFFER_MAX = 1920;
     MOZ_ASSERT((track_rate_/100)*sizeof(uint16_t) * 2 <= AUDIO_SAMPLE_BUFFER_MAX);