fix(ios): enable voice-processing I/O (echo cancellation) for voiceChat/videoChat session modes

[felixlabsco]

Problem

Setting AudioManager.setAudioSessionOptions({ iosMode: 'voiceChat', ... }) configures AVAudioSession routing and EQ, but does not enable echo cancellation — that requires setVoiceProcessingEnabled: on the engine's input node, which the library never calls.

Consequence for any full-duplex app (simultaneous AudioRecorder capture + playback, e.g. VoIP or realtime voice-agent clients): speaker output loops straight back into the microphone stream. Downstream VAD (local or server-side, e.g. speech-to-speech APIs with barge-in) detects the app's own playback as user speech, so the agent interrupts itself — in my testing this manifested as playback chopped into ~1s fragments as the remote VAD repeatedly fired on the echo. Developers setting voiceChat mode almost universally expect AEC to be part of the deal, which makes this a silent trap.

Fix

When the input node is materialized, sync Apple's voice-processing I/O (AEC + noise suppression + AGC) with the desired session mode: enabled for voiceChat/videoChat, disabled otherwise. No public API change.

Ordering constraints handled:

• runs before currentInputConnectionFormat is read — enabling voice processing changes the hardware format;
• only toggles while the engine is stopped (setVoiceProcessingEnabled: fails on a running engine);
• no-ops when the state already matches, and syncs both ways so leaving voice-chat mode restores the raw path.

Testing

Verified on device (iPhone, iOS 26): with iosCategory: 'playAndRecord', iosMode: 'voiceChat', streaming 16 kHz mic PCM via AudioRecorder.onAudioReady while playing 24 kHz PCM through an AudioBufferQueueSourceNode on the built-in speaker at full volume:

• before: remote VAD fired continuously on echo; playback self-interrupted every ~1s;
• after: zero echo-triggered VAD events over multi-minute sessions; deliberate talk-over (barge-in) still detected correctly.

Known inherent side effect of Apple's voice processing (documented behavior, not introduced by this PR beyond enabling the unit): slight output level reduction on the speaker route.

I first shipped this as a pnpm patch against 0.13.2 (enable-only variant) and it's been solid; this PR is the upstreamed, mode-synced version. Happy to rework it as an explicit opt-in option (e.g. a SessionOptions flag) if you'd rather not tie it to the session mode — mode-implied seemed like the least-surprise default since voiceChat without AEC is rarely what anyone wants.

[mdydek]

Looks solid, need probably a one or two days to wrap my head around it and decide on the approach, but its a good baseline

[mdydek]

I applied your approach on top of the #1210, because it was closer to my vision. Thank you for the contribution nevertheless, your work made my work much easier.

[felixlabsco]

Glad to hear it! Love the project.