Android: platform AEC doesn't cancel app's own playback unless the OUTPUT stream uses Usage::VoiceCommunication

[arekrgw]

Feature Request

When building a full-duplex voice assistant (mic capture + TTS playback at the same
time), enabling androidInputPreset: 'voiceCommunication' on the AudioRecorder alone
was not enough to get working acoustic echo cancellation on Android. Our STT kept
transcribing the assistant's own TTS playback.

The missing piece: the output stream (AudioPlayer.cpp) is opened without any
Usage, which defaults to USAGE_MEDIA. On many devices, the platform AEC chain
engaged by the VOICE_COMMUNICATION input preset only references the
voice-communication output path and media-stream playback leaks into the microphone
uncancelled.

Adding two lines to the output stream builder fixed it completely:

// AudioPlayer.cpp — openAudioStream()
builder.setSharingMode(SharingMode::Exclusive)
    ->setFormat(AudioFormat::Float)
    // ...existing config...
    ->setUsage(Usage::VoiceCommunication)   // ← added
    ->setContentType(ContentType::Speech)  // ← added
    // ...

Environment

• react-native-audio-api 1.0.0-nightly-f8b57cd-20260825
• React Native 0.86.2 / Expo SDK 57, physical Android device (loudspeaker, no headphones)
• Capture: AudioRecorder with { androidInputPreset: 'voiceCommunication', iosVoiceProcessing: true }
• Playback: AudioContext playing TTS PCM buffers
• Use case: streaming STT (websocket ASR) while speaking TTS replies (barge-in style interaction)

Symptom before the patch

• Speech-to-text transcribed the assistant's own spoken replies as if they were user input
• Token-overlap heuristics and LLM-based intent filtering reduced but never eliminated
  the self-echo turns
• Verified the input preset reaches Oboe correctly (setInputPreset(VoiceCommunication)
  present in openAudioStream()), so the gap was on the output/reference side

After the patch

• Self-echo eliminated at DSP level (zero phantom transcripts during playback)
• Real barge-in (user talking over the reply) works reliably
• No audio-routing regressions on our test device: playback still comes out of the
  loudspeaker as before

Suggestions

1. Library behavior: consider applying Usage::VoiceCommunication (+ ContentType::Speech)
   to the output stream automatically when an active recorder uses
   androidInputPreset: 'voiceCommunication' or expose an option for it.
2. Docs (next → AudioRecorder): the full-duplex section recommends
   androidInputPreset: 'voiceCommunication' + iosVoiceProcessing: true, but doesn't
   mention that the playback side must also sit on the voice-communication path for
   Android AEC to reference it.

[closetcaiman]

Hi @arekrgw, thanks for pointing this out. I need to think through the best approach, but we will address this for sure.