fix(lotus-denoise): reliability — never-silent watchdog, resume timeout, cache + activation

Track-A robustness fixes from the engine review; no quality/model changes.

- H1: auto-resume the AudioContext on `statechange` if it suspends mid-call
  (mobile backgrounding / audio interruption). Previously the dest node emitted
  digital silence with no recovery — a silent mute of the sender.
- H2: `resumeCtx()` races `resume()` against a timeout. A suspended context can
  only resume on a user gesture; the action can arrive via postMessage, so a
  bare `await resume()` inside LiveKit's track-change lock could hang and
  deadlock all later mute/unmute/device-switch. Now it proceeds and the H1
  watcher heals it.
- M1: don't cache a REJECTED wasm fetch — a transient blip during a reconnect
  used to permanently disable denoise for the session. Evict on failure.
- M2: activate denoise off `allConnections$` (local participant's connections)
  instead of `livekitRoomItems$`, which excludes the local participant and only
  surfaces rooms with a remote member — so denoise now also runs when you're
  alone and no longer couples to a remote-render concern.
- Context lifecycle: `closeContext()` removes the state watcher before closing;
  `ensureContext()` closes a half-initialised context on any failure (no leak).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
This commit is contained in:
Lotus CI
2026-06-30 23:27:44 -04:00
co-authored by Claude Opus 4.8
parent a4309f3e0c
commit 9a4e9bf7da
2 changed files with 79 additions and 26 deletions
+6 -2
View File
@@ -101,8 +101,12 @@ export function startLotusDenoise(vm: CallViewModel): () => void {
const roomListeners = new Map<LivekitRoom, () => void>();
let rooms: LivekitRoom[] = [];
const sub = vm.livekitRoomItems$.subscribe((items) => {
const next = items.map((i) => i.livekitRoom);
// Drive activation off the LOCAL participant's connection(s), not
// `livekitRoomItems$` — that stream excludes the local participant and only
// surfaces rooms with ≥1 remote member, so it wouldn't denoise you while
// you're alone and is a fragile coupling to a remote-render concern.
const sub = vm.allConnections$.subscribe((data) => {
const next = data.getConnections().map((c) => c.livekitRoom);
rooms = next;
for (const [room, off] of roomListeners) {
if (!next.includes(room)) {
+73 -24
View File
@@ -64,18 +64,41 @@ const sampleRateFor = (model: LotusDenoiseModel): number =>
// Cache fetched wasm per URL so a reconnect/device-switch doesn't re-download.
const wasmCache = new Map<string, Promise<ArrayBuffer>>();
function fetchWasm(url: string): Promise<ArrayBuffer> {
async function fetchWasmUncached(url: string): Promise<ArrayBuffer> {
const r = await fetch(url);
if (!r.ok) throw new Error(`denoise wasm ${url} -> ${r.status}`);
return r.arrayBuffer();
}
async function fetchWasm(url: string): Promise<ArrayBuffer> {
let p = wasmCache.get(url);
if (!p) {
p = fetch(url).then((r) => {
if (!r.ok) throw new Error(`denoise wasm ${url} -> ${r.status}`);
return r.arrayBuffer();
});
p = fetchWasmUncached(url);
// Never cache a REJECTED fetch: a transient failure (e.g. a blip during a
// reconnect) must not permanently disable denoise for the whole session.
// Evict on failure so the next restart/device-switch retries.
void p.catch(() => wasmCache.delete(url));
wasmCache.set(url, p);
}
return p;
}
/**
* Resume an AudioContext, but never block indefinitely. A suspended context can
* only resume after a user gesture; the denoise action can arrive via host
* postMessage (no gesture), so `resume()` may stay pending forever. Since this
* runs inside LiveKit's per-track change lock, a hung resume() would deadlock
* every later mute/unmute/device-switch. Race it against a timeout and proceed
* either way — the processor's `statechange` watcher resumes it once a gesture
* lands, and a still-suspended context degrades to (temporary) silence that the
* watcher heals, not a hang.
*/
async function resumeCtx(ctx: AudioContext, timeoutMs = 3_000): Promise<void> {
await Promise.race([
ctx.resume().catch(() => undefined),
new Promise<void>((resolve) => setTimeout(resolve, timeoutMs)),
]);
}
function supportsSimd(): boolean {
try {
// Minimal SIMD module (v128) — validates only where SIMD is supported.
@@ -123,6 +146,7 @@ export class LotusDenoiseProcessor
private ctx?: AudioContext;
private graph?: Graph;
private ctxStateHandler?: () => void;
public constructor(private readonly config: LotusDenoiseConfig) {}
@@ -152,38 +176,63 @@ export class LotusDenoiseProcessor
this.disposeGraph(this.graph);
this.graph = undefined;
this.processedTrack = undefined;
if (this.ctx && this.ctx.state !== "closed")
await this.ctx.close().catch(() => undefined);
await this.closeContext();
}
/** Remove the state watcher and close the owned context, if any. */
private async closeContext(): Promise<void> {
const ctx = this.ctx;
if (!ctx) return;
if (this.ctxStateHandler) {
ctx.removeEventListener("statechange", this.ctxStateHandler);
this.ctxStateHandler = undefined;
}
this.ctx = undefined;
if (ctx.state !== "closed") await ctx.close().catch(() => undefined);
}
/** Create (once) the model-rate context + register the flat worklet modules. */
private async ensureContext(): Promise<void> {
const rate = sampleRateFor(this.config.model);
if (this.ctx && this.ctx.state !== "closed" && this.ctx.sampleRate === rate) {
if (this.ctx.state === "suspended") await this.ctx.resume();
if (this.ctx.state === "suspended") await resumeCtx(this.ctx);
return;
}
if (this.ctx && this.ctx.state !== "closed")
await this.ctx.close().catch(() => undefined);
await this.closeContext();
const ctx = new AudioContext({ sampleRate: rate });
if (ctx.sampleRate !== rate) {
try {
if (ctx.sampleRate !== rate)
throw new Error(`denoise: got ${ctx.sampleRate}Hz, need ${rate}Hz`);
// Auto-resume if the OS/browser suspends the context mid-call (mobile
// backgrounding, audio interruption): the dest node otherwise emits
// silence with no recovery. Only resume while a graph is live.
const onStateChange = (): void => {
if (ctx.state === "suspended" && this.graph)
void ctx.resume().catch(() => undefined);
};
ctx.addEventListener("statechange", onStateChange);
// Flat models register via addModule here; DTLN/DeepFilterNet bring their
// own processor via the dynamic-imported helper (see buildMlNode).
if (this.config.model === "rnnoise" || this.config.model === "speex")
await ctx.audioWorklet.addModule(
this.config.assetBase + FLAT[this.config.model].script,
);
if (this.config.gate)
await ctx.audioWorklet.addModule(this.config.assetBase + GATE.script);
// The action can arrive via host postMessage, not a gesture in this
// iframe, so the context can start suspended — resume without hanging.
if (ctx.state === "suspended") await resumeCtx(ctx);
this.ctx = ctx;
this.ctxStateHandler = onStateChange;
} catch (e) {
// Don't leak a half-initialised context on any failure path.
await ctx.close().catch(() => undefined);
throw new Error(`denoise: got ${ctx.sampleRate}Hz, need ${rate}Hz`);
throw e;
}
// Flat models register via addModule here; DTLN/DeepFilterNet bring their
// own processor via the dynamic-imported helper (see buildMlNode).
if (this.config.model === "rnnoise" || this.config.model === "speex")
await ctx.audioWorklet.addModule(
this.config.assetBase + FLAT[this.config.model].script,
);
if (this.config.gate)
await ctx.audioWorklet.addModule(this.config.assetBase + GATE.script);
// The action arrives via host postMessage, not a gesture in this iframe, so
// the context can start suspended.
if (ctx.state === "suspended") await ctx.resume();
this.ctx = ctx;
}
private async buildGraph(track: MediaStreamTrack): Promise<Graph> {