Integrate DeepFilterNet 3 (deepfilternet3-noise-filter@1.2.1) as a new
client-side denoise model id 'deepfilternet', mirroring the DTLN pattern.
The npm package ships only an ESM whose AudioWorklet processor + wasm-bindgen
glue are inlined as a string (loaded via a Blob URL — no CDN for the worklet).
Its only runtime fetches are a single-threaded df_bg.wasm and an ONNX model
tarball, which previously loaded from an external CDN. We now VENDOR both
(build/denoise-vendor/deepfilternet/v2/...) and self-host them under
denoise/deepfilternet/, overriding the package's cdnUrl so nothing hits the
upstream CDN — keeping it self-hosted / Tauri-CSP safe.
The wasm is single-threaded (no SharedArrayBuffer / atomics / imported shared
memory), so it needs no COOP/COEP cross-origin isolation and runs fine in EC's
non-isolated iframe. Runs at 48 kHz fullband. Any init/runtime failure falls
back to the raw mic, like the other models.
- vite.config.js: copy ESM + vendored wasm/model into the EC denoise dir with a
required-asset guard that aborts the build if any entry is missing.
- build/lotus-denoise.js: 'deepfilternet' branch — dynamic-import the ESM, build
a DeepFilterNet3Core pointed at the self-hosted base, await init, return the
worklet node; 48 kHz; raw-mic fail-safe preserved.
- denoisePipeline.ts: 'deepfilternet' branch for the in-app tester + sampleRate.
- settings.ts: add 'deepfilternet' to DenoiseModelId + getSettings whitelist.
- lotusDenoiseUtils.ts: add the comparison-chart row.
- General.tsx: add the "DeepFilterNet 3 (beta)" dropdown option.
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 prior DTLN attempt (89a2321d) broke the build (missing dep, wrong
`cinny/` asset paths) and typecheck (`'dtln'` not in DenoiseModelId), and was
wired against an API the package doesn't expose. @workadventure/noise-
suppression is not a flat AudioWorklet — it's a self-contained ES module whose
processor name is `workadventure-noise-suppression` and which resolves its own
LiteRT WASM + TFLite models via import.meta.url. Driving it by hand-rolled
addModule + processorOptions cannot work.
- Re-add @workadventure/noise-suppression@0.0.4 (package.json + lockfile).
- vite: copy the package's whole dist/ tree intact to
denoise/workadventure/ (preserving assets/ + vendor/litert) so import.meta
resolution works at runtime; fail the build if the entry module is missing.
- shim: for the DTLN model, dynamic-import denoise/workadventure/audio-worklet
.js and use createNoiseSuppressionAudioWorklet(ctx, { bypassUntilReady })
to build the node; RNNoise/Speex keep their direct flat-worklet path. Async
init errors are logged + reported and fall back to the raw mic.
- Restore 'dtln' in DenoiseModelId (+ settings coercion), the model chart, and
the settings dropdown, labelled "(beta)".
DTLN builds and is fully self-hosted, but its in-call audio is UNVERIFIED in
this environment — needs a real-call test. DeepFilterNet stays excluded (CDN
asset loading, incompatible with self-hosting / Tauri CSP).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Audit/repair of the multi-model denoise work so it actually builds and only
exposes working, self-hosted models.
- Complete the DTLN/DFN3 revert: uninstall @workadventure/noise-suppression
and deepfilternet3-noise-filter (package.json + lockfile), drop the unused
DTLN asset-copy block from vite.config.js (was shipping ~2MB of unused
tflite/wasm), and narrow DenoiseModelId to the bundled models (rnnoise,
speex). Coerce any retired persisted model value back to the default.
- Fix General.tsx CI typecheck failures introduced by the denoise UI: restore
three imports the rewrite deleted (useDateFormatItems, SequenceCardStyle,
useTauriUpdater), add the missing denoise/sound imports, and correct
hallucinated Folds props (Text has no variant/bold; Box uses
alignItems/justifyContent). tsc now passes with 0 errors.
- Harden the vite denoise plugin: required RNNoise/Speex/gate assets and the
shim now fail the build loudly if missing (instead of a silent warn that
shipped a broken ML feature), and the index.html shim injection is verified.
- CI: move the cinny-desktop submodule bump into ci.yml as a `trigger-desktop`
job gated on `needs: build`, and delete the standalone trigger-desktop.yml.
A failing push no longer kicks off the slow Tauri builds in parallel.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Remove non-functional DTLN and DFN3 dependencies and UI options.
- Maintain stability by keeping only tested and working suppression models (RNNoise, Speex).
- Verified that build passes and all assets are correctly bundled.
- Verified package layouts and integration paths for @workadventure/noise-suppression (v0.0.4) and deepfilternet3-noise-filter (v1.2.1).
- Updated build configuration to correctly copy WASM, TFLite, and ONNX assets.
- Integrated DTLN and DeepFilterNet initialization logic into the audio shim.
- Enabled all four models (RNNoise, Speex, DTLN, DFN3) in Settings UI.
Implement a flexible, multi-model noise suppression pipeline for Element Call/LiveKit integration:
- ML Engines: Added support for RNNoise, Speex, DTLN, and DeepFilterNet 3 models.
- Pipeline Architecture: Implemented modular audio processing in lotus-denoise.js, supporting 'Series Suppression' (running browser-native NSNet2 before ML) and a hardware-style Noise Gate.
- UI & UX Enhancements:
- Settings UI: Added model comparison chart with CPU/Quality metadata.
- Tuning: Added Live Microphone Meter for calibrating Noise Gate thresholds.
- Reporting: Added LotusToast system to alert users when ML suppression fails or falls back to raw input.
- Robustness & Quality:
- Capture Fidelity: Removed forced 48kHz capture constraints to allow native-rate capture (solving static issues with high-end audio interfaces).
- Performance: Added WASM SIMD detection with transparent fallback.
- Capability Detection: Added browser feature detection to disable unsupported ML modes.
- Build Integration: Updated Vite config to self-host all model WASM/tflite assets in /denoise/ directory.
ML noise suppression produced loud static on real calls. RNNoise requires
mono 48kHz float input; feeding it stereo or wrong-rate data is the classic
cause of that static. Harden the shim:
- request mono (channelCount:1) + 48kHz capture
- run a 48kHz AudioContext and BAIL to the raw mic if the browser won't
give a true 48kHz context (wrong-rate data -> static)
- force the worklet node to explicit mono in/out
- use the non-SIMD rnnoise.wasm (SIMD build artifacts on some GPUs)
- share one AudioContext across captures
Also fix the two CI-blocking eslint errors (unused vars in UrlPreviewCard
and useLocalMessageSearch) and apply repo-wide prettier formatting so
check:eslint and check:prettier pass.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Replace the boolean call noise-suppression setting with a 3-way control
(Off / Browser-native / ML beta) in Settings -> General -> Calls.
- Off: noiseSuppression=false to Element Call
- Browser-native: EC's built-in WebRTC suppressor (prior default)
- ML (beta): on-device RNNoise (@sapphi-red/web-noise-suppressor)
Element Call captures the mic inside its iframe and publishes to LiveKit,
so the host can't reach that track; LiveKit's Krisp filter is Cloud-only
(we self-host the SFU) and EC's own RNNoise PR #3892 is unmerged. The ML
tier instead injects a same-origin pre-init shim into the vendored EC
index.html (build/lotus-denoise.js, wired by the lotusDenoise vite plugin)
that patches getUserMedia and routes the captured mic through an RNNoise
AudioWorklet before LiveKit sees it -- the same post-capture pipeline as
#3892, with no EC fork/AGPL/rebase burden. Falls back to the raw mic if
setup fails; keeps echoCancellation/AGC on the raw capture.
- settings.ts: callNoiseSuppression -> 'off'|'browser'|'ml' + legacy
boolean migration (true->browser, false->off)
- CallEmbed/useCallEmbed: tier maps to noiseSuppression param and appends
lotusDenoise=ml (native suppressor off in ML mode)
- vite.config.js: copy RNNoise worklet/wasm + shim into the EC bundle and
inject the shim <script> before EC's module entry
- docs: LOTUS_FEATURES.md, LOTUS_TODO.md (P5-30 done)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
jared