feat(denoise): add self-hosted DeepFilterNet 3 ML noise-suppression model

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>
This commit is contained in:
2026-06-17 19:57:08 -04:00
parent abb7f743b8
commit 04b56ffacd
10 changed files with 141 additions and 14 deletions
@@ -1358,6 +1358,7 @@ function Calls() {
{ value: 'rnnoise', label: 'RNNoise' },
{ value: 'speex', label: 'Speex (Legacy)' },
{ value: 'dtln', label: 'DTLN (beta)' },
{ value: 'deepfilternet', label: 'DeepFilterNet 3 (beta)' },
]}
/>
}
+10 -7
View File
@@ -14,10 +14,12 @@ export type MessageSpacing = '0' | '100' | '200' | '300' | '400' | '500';
// - 'browser' : WebRTC built-in suppression (Element Call noiseSuppression param)
// - 'ml' : client-side RNNoise ML suppression (Lotus denoise shim)
export type NoiseSuppressionMode = 'off' | 'browser' | 'ml';
// Self-hostable, build-bundled ML models. DeepFilterNet remains excluded — it
// loads its runtime/models from external CDNs, which breaks the self-hosted /
// Tauri-CSP strategy (see LOTUS_DENOISE_ENGINEERING_REVIEW.md).
export type DenoiseModelId = 'rnnoise' | 'speex' | 'dtln';
// Self-hostable, build-bundled ML models. DeepFilterNet 3 is included via
// deepfilternet3-noise-filter with its df_bg.wasm + ONNX model VENDORED and
// self-hosted (its cdnUrl is overridden), so it no longer depends on an external
// CDN. Its wasm is single-threaded, so no COOP/COEP cross-origin isolation is
// required (see LOTUS_DENOISE_ENGINEERING_REVIEW.md).
export type DenoiseModelId = 'rnnoise' | 'speex' | 'dtln' | 'deepfilternet';
export type ChatBackground =
| 'none'
| 'blueprint'
@@ -260,12 +262,13 @@ export const getSettings = (): Settings => {
? 'browser'
: 'off'
: (saved.callNoiseSuppression ?? defaultSettings.callNoiseSuppression),
// Coerce any retired/unknown persisted model (e.g. 'dtln', 'deepfilternet'
// from earlier beta builds) back to the default working model.
// Coerce any retired/unknown persisted model back to the default working
// model; only whitelisted ids pass through.
callDenoiseModel:
saved.callDenoiseModel === 'rnnoise' ||
saved.callDenoiseModel === 'speex' ||
saved.callDenoiseModel === 'dtln'
saved.callDenoiseModel === 'dtln' ||
saved.callDenoiseModel === 'deepfilternet'
? saved.callDenoiseModel
: defaultSettings.callDenoiseModel,
composerToolbarButtons: {
+30 -4
View File
@@ -14,10 +14,10 @@ import { DenoiseModelId } from '../state/settings';
const BASE = `${import.meta.env.BASE_URL.replace(/\/+$/, '')}/public/element-call/denoise/`;
/**
* Required AudioContext sample rate per model. RNNoise/Speex (sapphi) assume
* 48 kHz. DTLN (@workadventure) targets 16 kHz and does NOT resample internally
* — running it at 48 kHz produces robotic/choppy/quiet output, so its whole
* graph must run in a 16 kHz context.
* Required AudioContext sample rate per model. RNNoise/Speex (sapphi) and
* DeepFilterNet 3 are 48 kHz. DTLN (@workadventure) targets 16 kHz and does NOT
* resample internally — running it at 48 kHz produces robotic/choppy/quiet
* output, so its whole graph must run in a 16 kHz context.
*/
export const sampleRateFor = (model: DenoiseModelId): number => (model === 'dtln' ? 16000 : 48000);
@@ -75,6 +75,32 @@ export async function buildModelNode(
const handle = await mod.createNoiseSuppressionAudioWorklet(ctx, { bypassUntilReady: true });
return { node: handle.node, dispose: () => handle.dispose() };
}
if (model === 'deepfilternet') {
// deepfilternet3-noise-filter ESM: inlines its worklet/wasm-bindgen glue and
// fetches only df_bg.wasm + the ONNX model, which we self-host under
// deepfilternet/v2/... Override its cdnUrl to our absolute base so nothing
// hits the upstream CDN. DeepFilterNet3Core builds the worklet node directly.
const dfnBase = new URL(`${BASE}deepfilternet`, window.location.href).href;
// eslint-disable-next-line @typescript-eslint/no-explicit-any
const mod: any = await import(/* @vite-ignore */ `${BASE}deepfilternet/index.esm.js`);
const core = new mod.DeepFilterNet3Core({
sampleRate: sampleRateFor(model),
noiseReductionLevel: 80,
assetConfig: { cdnUrl: dfnBase },
});
await core.initialize();
const node: AudioWorkletNode = await core.createAudioWorkletNode(ctx);
return {
node,
dispose: () => {
try {
core.destroy();
} catch {
/* noop */
}
},
};
}
const cfg = SAPPHI[model];
const [, wasmBinary] = await Promise.all([addModuleOnce(ctx, cfg.script), fetchWasm(cfg.wasm)]);
const node = new AudioWorkletNode(ctx, cfg.proc, {
+9
View File
@@ -40,6 +40,15 @@ export const DENOISE_MODELS: DenoiseModel[] = [
transients: 'Excellent',
voiceQuality: 'High',
},
{
id: 'deepfilternet',
name: 'DeepFilterNet 3 (beta)',
description: 'Studio-grade deep-learning model (48 kHz, ONNX). Best quality; highest CPU.',
cpuUsage: '25-50%',
binarySize: '~18 MB',
transients: 'Excellent',
voiceQuality: 'Very High',
},
];
export const isMLDenoiseSupported = (): boolean => {