The echo gate architecture bypassed the serialized mic management chain
and introduced race conditions. Chrome's AEC handles echo properly for
both headphone and speaker users without custom intervention.
LiveKit's server-side VAD has a high, non-configurable threshold that misses
conversational speech. Replace it with a SpeakingDetector singleton that uses
Web Audio AnalyserNodes to read actual RMS levels per participant (50ms poll,
0.008 threshold, 250ms hysteresis hold).
tsc was emitting compiled .js files directly into packages/web/src/
alongside the .tsx source files because noEmit was not set. These
artifacts were never used — Vite compiles from .tsx source directly.
- Add noEmit: true to packages/web/tsconfig.json (tsc = type-check only)
- Delete all 62 orphaned .js files from src/ (-6,129 lines)
- Add packages/web/src/**/*.js to .gitignore as safeguard
RNNoise worklet outputs mono (1ch) which played only in the left ear.
Replace unreliable automatic up-mixing (channelCount/channelCountMode on
inputGain) with a ChannelMergerNode that duplicates the mono signal to
both L and R channels — guaranteed stereo by the Web Audio spec.
Inject @sapphi-red/web-noise-suppressor into AudioManager's input pipeline
as a toggleable AudioWorkletNode. The worklet is loaded lazily on first
enable, then kept alive — toggling bypasses by rewiring the graph without
destroying the WASM instance. Browser NS is forced off when RNNoise is
active to avoid double-processing. InputGain forced to stereo up-mix to
prevent mono-left-only output from the worklet. Also wires Phase 2 output
device routing (setSinkId) into AppLayout/ChannelSidebar.
Chrome's AEC uses getDisplayMedia audio as a reference signal and
aggressively ducks the microphone even when headphones are used.
This adds a screenShareActive flag to AudioManager that forces
echoCancellation off during screen share, plus Chromium-specific
goog* constraints as belt-and-suspenders.
Disable AGC by default to prevent Chrome from crushing mic sensitivity
when stream audio is playing. Add user-facing toggles for echo
cancellation, noise suppression, and auto gain control. Decouple voice
and stream audio by routing through ctx.destination instead of shared
compressor. Track mic stream generation to re-publish when settings change.
Add MutationObserver to neutralize LiveKit's re-attached <audio> elements,
mark our keep-alive elements with data-opencord, detach tracks on unsubscribe,
remove dangerous blanket .play(), and soften compressor to transparent limiter.
Streams now appear as separate tiles in the voice grid alongside the
user's camera/avatar tile, matching Discord's model. Each stream tile
has independent volume, mute, watch/unwatch controls, quality badges,
and stream attenuation that ducks audio when someone speaks.
Jannis Braun