Edit history

Feature Contribution

I've implemented three new audio processing features for the voice system and would like to offer them upstream. The implementation is complete and tested on my self-hosted instance. Branch: cproudlock:echowire-refactor (diff)

I wasn't able to create a PR directly — if fork PRs are enabled, I'm happy to resubmit as a proper PR.

---

Summary

Adds three new audio processing features using @sapphi-red/web-noise-suppressor and the Web Audio API:

• ML Noise Suppression (RNNoise) — WASM-based neural network that filters background noise (fans, typing, traffic). Uses AudioWorklet for off-main-thread processing at 48kHz with ~13ms latency.
• Noise Gate — Mutes the microphone when audio falls below a configurable dB threshold (-100 to 0). Uses hysteresis (separate open/close thresholds) to prevent rapid toggling.
• Compressor — DynamicsCompressorNode that evens out loud and quiet audio for more consistent voice levels.

Processing Chain

When all three are disabled, no processor is attached — zero overhead.

Why not @livekit/krisp-noise-filter?

Krisp requires LiveKit Cloud and doesn't work with self-hosted LiveKit. @sapphi-red/web-noise-suppressor provides ready-made AudioWorklet nodes for RNNoise with WASM SIMD auto-detection, making it a drop-in solution for any deployment.

Files Changed

Settings Added

Implementation Details

• Uses LiveKit's TrackProcessor<Track.Kind.Audio> interface for clean integration
• AudioWorklet files + WASM binaries copied to dist/audio-worklets/ via rspack CopyPlugin
• Settings persisted via MobX persistence (IndexedDB), changes apply immediately mid-call
• Processor rebuilds graph on settings change; fully tears down when all processing disabled
• RNNoise uses SIMD auto-detection for best performance where available
• Noise gate uses 90ms hold time and -10dB hysteresis gap to prevent chatter

Feature Contribution

I've implemented three new audio processing features for the voice system and would like to offer them upstream. The implementation is complete and tested on my self-hosted instance. Branch: cproudlock:echowire-refactor (diff)

I wasn't able to create a PR directly — if fork PRs are enabled, I'm happy to resubmit as a proper PR.

---

Summary

Adds three new audio processing features using @sapphi-red/web-noise-suppressor and the Web Audio API:

• ML Noise Suppression (RNNoise) — WASM-based neural network that filters background noise (fans, typing, traffic). Uses AudioWorklet for off-main-thread processing at 48kHz with ~13ms latency.
• Noise Gate — Mutes the microphone when audio falls below a configurable dB threshold (-100 to 0). Uses hysteresis (separate open/close thresholds) to prevent rapid toggling.
• Compressor — DynamicsCompressorNode that evens out loud and quiet audio for more consistent voice levels.

Processing Chain

When all three are disabled, no processor is attached — zero overhead.

Why not @livekit/krisp-noise-filter?

Krisp requires LiveKit Cloud and doesn't work with self-hosted LiveKit. @sapphi-red/web-noise-suppressor provides ready-made AudioWorklet nodes for RNNoise with WASM SIMD auto-detection, making it a drop-in solution for any deployment.

Files Changed

Settings Added

Implementation Details

• Uses LiveKit's TrackProcessor<Track.Kind.Audio> interface for clean integration
• AudioWorklet files + WASM binaries copied to dist/audio-worklets/ via rspack CopyPlugin
• Settings persisted via MobX persistence (IndexedDB), changes apply immediately mid-call
• Processor rebuilds graph on settings change; fully tears down when all processing disabled
• RNNoise uses SIMD auto-detection for best performance where available
• Noise gate uses 90ms hold time and -10dB hysteresis gap to prevent chatter