- Recorded clips often decode as stereo (Chrome Opus/WebM). The models
only process channel 0, and the default channel mode ignores
channelCount, so RNNoise/Speex played the cleaned audio in the LEFT EAR
ONLY. Downmix to mono before the models (as Lotus Call does, #24) and
pin the RNNoise/Speex/gate nodes to explicit mono. Verified in Chromium
with a stereo speech clip through RNNoise: before L 16.1 / R 0.0, after
L 16.0 / R 15.9.
- DeepFilterNet ran at attenuation limit 80 here vs 60 in calls; both
muted the background outright, which sounds like the mic cutting out
at word ends. Use 25, matching element-call#65.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017PCb59uY3UKqDqt7B6hzd5
Settings → Calls A/B denoise tester leaked audio resources:
- play() built a denoise model node (DeepFilterNet/DTLN worker/WASM) + optional
gate but stopPlayback only closed the AudioContext, never disposing them —
each A/B playback-through-a-model leaked a worker. stopPlayback now mirrors
stopLive (gate.disconnect → model.dispose → node.disconnect).
- A generation token (playGenRef, bumped by stopPlayback) makes play() discard
what it built if superseded during the async WASM/worklet load — closing the
same leak in the rapid-Play-click race, the Stop-during-load case, and the
unmount-during-load case, and stopping a superseded rejection from tearing
down the winning playback.
- A mountedRef guards the getUserMedia paths (startLive/startRecord) so closing
Settings during the mic permission prompt doesn't create untracked
resources / setState-after-unmount; its effect sets true on mount (not only
false on cleanup) so it survives a StrictMode/Activity same-fiber remount.
Bug-hunt findings from LOTUS_TODO. Three review passes: the first two confirmed
the base fixes and surfaced the concurrent-load leak + StrictMode fragility; a
third traced all six play() interleavings of the generation token. Gate-green
(tsc, eslint, prettier, 914 tests, build).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Move the model comparison out of the always-visible Noise Suppression
description and into the ML-only sub-settings. Add a compact info card
for the selected model (CPU / voice quality / transients / download) plus
a collapsible 4-model comparison. Group ML sub-settings into Model,
Enhancements, and Test & calibrate sections with clear labels and
separators. Fix invented --lt-border-color token and hardcoded
rgba background to real TDS tokens. Build the model dropdown and
DenoiseTester labels/compare buttons from DENOISE_MODELS so
DeepFilterNet 3 is handled correctly.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Two issues found from real testing of the in-app tester:
1. Raw ≈ RNNoise ≈ Speex sounded identical in Record & compare because the clip
was captured with browser noise suppression ON (the user's native-NS
setting), so "Raw" was already cleaned and the models had nothing left to
remove. Record & compare now captures fully raw audio (noiseSuppression /
AGC / echoCancellation off) so each model's effect on real noise is audible.
(Friends still heard differences in calls — the models work; the test was
feeding them pre-cleaned audio.)
2. DTLN was robotic/choppy/quiet because @workadventure/noise-suppression
targets 16 kHz (AUDIO_CONFIG.sampleRate) and does NOT resample internally,
while we ran it at 48 kHz. Run DTLN's whole graph in a 16 kHz context:
- denoisePipeline: add sampleRateFor(model) (16k for dtln, 48k otherwise);
tester live-monitor + playback contexts use it (bufferSource resamples the
48k clip down for DTLN).
- shim (build/lotus-denoise.js): SAMPLE_RATE is now model-aware, so DTLN is
correct in real calls too (it was previously broken at 48 kHz). The 16 kHz
processed track is still published to LiveKit (WebRTC/Opus resamples).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The previous "Test Microphone" meter only showed a raw 0-100% level bar — it
never ran the gate or any model, and its scale wasn't dBFS, so it couldn't tell
you which threshold to pick or let you hear the models solo. Replace it with a
real tester that reuses the shipped worklets (/public/element-call/denoise/) in
a main-app AudioContext, mirroring the call pipeline (source -> gate -> model).
- denoisePipeline.ts: shared loader for the RNNoise/Speex flat worklets and the
DTLN @workadventure helper, the noise gate, and a dBFS RMS meter helper.
- DenoiseTester.tsx:
- Live monitor: hear yourself through the selected model (+gate) in real time
(headphones) with In/Out dBFS meters and a threshold marker on the In meter
so the gate value is meaningful to calibrate.
- Record & compare: capture a short clip, then A/B the same audio Raw vs
RNNoise vs Speex vs DTLN.
- Wire it into the ML settings block; remove the old raw-only MicMeter. Use real
TDS tokens (--accent-*, --border-color, --bg-card) instead of the invented
--lt-* names + hardcoded hex the old meter used.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>