Denoise deep-review (CRITICAL): restart() read opts.audioContext, which LiveKit does NOT pass on restart — so reconnect (the A7 scenario) and mic device-switch threw after stopping the old track, leaving the mic SILENT (A7 reintroduced). Fix: - Processor owns a dedicated 48kHz AudioContext (sapphi worklets require 48kHz; H1), reused across restart, closed on destroy. - restart() never throws and never leaves a stopped track on the sender: builds the new graph first, then disposes the old; on failure degrades to RAW mic audio rather than silence. - Cache wasm per URL (no re-fetch each reconnect); gate threshold default -45 and accept an explicit 0 (M2); document the cross-repo asset contract. Protocol audit: - Non-silent warning when an unsupported denoise model (dtln/deepfilternet) is requested instead of silent rnnoise fallback (F3). - Correct the call_state enum comment (immediate error-reply, not 10s) (F2). Build/CI audit: - Stamp VITE_APP_VERSION in CI; document the vX.Y.Z-lotus.N version scheme. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
135 lines
4.5 KiB
TypeScript
135 lines
4.5 KiB
TypeScript
/*
|
|
Copyright 2026 Lotus Guild
|
|
|
|
SPDX-License-Identifier: AGPL-3.0-only OR LicenseRef-Element-Commercial
|
|
Please see LICENSE in the repository root for full details.
|
|
*/
|
|
|
|
import {
|
|
type LocalAudioTrack,
|
|
ParticipantEvent,
|
|
type Room as LivekitRoom,
|
|
Track,
|
|
} from "livekit-client";
|
|
import { logger } from "matrix-js-sdk/lib/logger";
|
|
|
|
import { type CallViewModel } from "../state/CallViewModel/CallViewModel";
|
|
import { lotusFlag, lotusParam } from "./lotusWidget";
|
|
import {
|
|
type LotusDenoiseConfig,
|
|
LotusDenoiseProcessor,
|
|
} from "./lotusDenoiseProcessor";
|
|
|
|
/**
|
|
* Apply Lotus ML noise suppression to the local mic track as a first-class
|
|
* Element Call audio TrackProcessor (#1), replacing the host's build-time
|
|
* `getUserMedia` monkeypatch.
|
|
*
|
|
* Opt-in via `lotusDenoise=ml` (+ `lotusModel`, `lotusGate`,
|
|
* `lotusGateThreshold`, `lotusDenoiseBase`). Because the processor is attached
|
|
* to the mic publication and re-attached on every (re)publish, denoise
|
|
* survives EC's mid-call reconnect — fixing the A7 "mic dead after reconnect"
|
|
* bug. Additive: no-op without the flag.
|
|
*/
|
|
/**
|
|
* Resolve the denoise asset base, forcing it to be SAME-ORIGIN. The base is fed
|
|
* to `audioWorklet.addModule()`, which executes the target as code in the
|
|
* worklet scope (processing the live mic) — so a cross-origin base from a
|
|
* crafted call link would be arbitrary code execution. Any non-same-origin or
|
|
* malformed value falls back to the bundled "./denoise/".
|
|
*/
|
|
function safeAssetBase(raw: string | null): string {
|
|
const fallback = "./denoise/";
|
|
if (!raw) return fallback;
|
|
try {
|
|
const u = new URL(raw, window.location.href);
|
|
if (u.origin !== window.location.origin) return fallback;
|
|
return u.href.endsWith("/") ? u.href : `${u.href}/`;
|
|
} catch {
|
|
return fallback;
|
|
}
|
|
}
|
|
|
|
export function startLotusDenoise(vm: CallViewModel): () => void {
|
|
if (lotusParam("lotusDenoise") !== "ml") return () => undefined;
|
|
|
|
const requestedModel = lotusParam("lotusModel");
|
|
// Only RNNoise/Speex are implemented in-source so far. Don't silently
|
|
// degrade an unsupported model (e.g. the host's dtln/deepfilternet) to
|
|
// rnnoise — log it so the mismatch is visible.
|
|
let model: LotusDenoiseConfig["model"] = "rnnoise";
|
|
if (requestedModel === "speex") model = "speex";
|
|
else if (requestedModel && requestedModel !== "rnnoise")
|
|
logger.warn(
|
|
`[lotus] denoise model "${requestedModel}" not implemented in-source; using rnnoise`,
|
|
);
|
|
|
|
const rawThreshold = lotusParam("lotusGateThreshold");
|
|
const config: LotusDenoiseConfig = {
|
|
model,
|
|
assetBase: safeAssetBase(lotusParam("lotusDenoiseBase")),
|
|
gate: lotusFlag("lotusGate"),
|
|
// Default -45 (matches the reference shim); accept an explicit 0 (don't
|
|
// coerce it away via `|| default`).
|
|
gateThreshold:
|
|
rawThreshold !== null && Number.isFinite(Number(rawThreshold))
|
|
? Number(rawThreshold)
|
|
: -45,
|
|
};
|
|
|
|
const micOf = (room: LivekitRoom): LocalAudioTrack | undefined =>
|
|
room.localParticipant.getTrackPublication(Track.Source.Microphone)
|
|
?.track as LocalAudioTrack | undefined;
|
|
|
|
const apply = (room: LivekitRoom): void => {
|
|
const mic = micOf(room);
|
|
if (mic && !mic.getProcessor()) {
|
|
void mic
|
|
.setProcessor(new LotusDenoiseProcessor(config))
|
|
.catch((e) => logger.warn("[lotus] denoise setProcessor failed", e));
|
|
}
|
|
};
|
|
|
|
const roomListeners = new Map<LivekitRoom, () => void>();
|
|
let rooms: LivekitRoom[] = [];
|
|
|
|
const sub = vm.livekitRoomItems$.subscribe((items) => {
|
|
const next = items.map((i) => i.livekitRoom);
|
|
rooms = next;
|
|
for (const [room, off] of roomListeners) {
|
|
if (!next.includes(room)) {
|
|
off();
|
|
roomListeners.delete(room);
|
|
}
|
|
}
|
|
for (const room of next) {
|
|
if (!roomListeners.has(room)) {
|
|
// Re-attach on every (re)publish — this is what makes denoise survive
|
|
// reconnects (A7), unlike the old getUserMedia patch.
|
|
const onPublished = (): void => apply(room);
|
|
room.localParticipant.on(
|
|
ParticipantEvent.LocalTrackPublished,
|
|
onPublished,
|
|
);
|
|
roomListeners.set(room, () =>
|
|
room.localParticipant.off(
|
|
ParticipantEvent.LocalTrackPublished,
|
|
onPublished,
|
|
),
|
|
);
|
|
apply(room);
|
|
}
|
|
}
|
|
});
|
|
|
|
return () => {
|
|
sub.unsubscribe();
|
|
for (const off of roomListeners.values()) off();
|
|
roomListeners.clear();
|
|
for (const room of rooms) {
|
|
const mic = micOf(room);
|
|
if (mic?.getProcessor()) void mic.stopProcessor();
|
|
}
|
|
};
|
|
}
|