feat(lotus): report speakingWhileMuted for the local participant in io.lotus.call_state (#37)

While the mic is muted LiveKit keeps the capture alive and merely disables the
published MediaStreamTrack, so VAD says nothing. lotusMutedSpeech taps a CLONE
of the published track (post-denoise when the in-source processor is active),
samples RMS at 10 Hz through an AnalyserNode and runs a hysteresis gate (300 ms
of voice on, 800 ms of quiet off, threshold ≈ −36 dBFS). The tap only exists
while a muted mic publication exists; the flag is set on the local entry of
io.lotus.call_state only and is never sent to other participants. Gate is
unit-tested; verified end-to-end: muted with a tone mic → true within 2.5 s,
silence → false, unmute → false.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PPmy3tPq869XDW4njjVaKA

src/lotus/lotusCallState.test.ts

@@ -53,7 +53,11 @@ function mockMember(id: string, userId: string): Member {
 }
 
 function mockVm(members: Member[]): CallViewModel {
-  return { userMedia$: of(members) } as unknown as CallViewModel;
+  return {
+    userMedia$: of(members),
+    // no livekit connections → the muted-speech tap never starts
+    allConnections$: of({ getConnections: () => [] }),
+  } as unknown as CallViewModel;
 }
 
 function participantsOf(call: number): unknown[] {

src/lotus/lotusCallState.ts

@@ -5,7 +5,7 @@ SPDX-License-Identifier: AGPL-3.0-only OR LicenseRef-Element-Commercial
 Please see LICENSE in the repository root for full details.
 */
 
-import { combineLatest, of, type Subscription } from "rxjs";
+import { combineLatest, of, startWith, type Subscription } from "rxjs";
 import {
   distinctUntilChanged,
   map,
@@ -16,6 +16,7 @@ import {
 import { type CallViewModel } from "../state/CallViewModel/CallViewModel";
 import { widget } from "../widget";
 import { LotusWidgetActions, lotusFlag, lotusSendToHost } from "./lotusWidget";
+import { observeSpeakingWhileMuted$ } from "./lotusMutedSpeech";
 
 interface ParticipantState {
   /** EC media id (`${userId}:${deviceId}`), stable per participant device. */
@@ -25,6 +26,11 @@ interface ParticipantState {
   speaking: boolean;
   audioEnabled: boolean;
   videoEnabled: boolean;
+  /**
+   * [lotus #37] LOCAL participant only: voice detected on the mic while it is
+   * muted (see lotusMutedSpeech.ts). Absent for remote participants.
+   */
+  speakingWhileMuted?: boolean;
 }
 
 /**
@@ -45,7 +51,8 @@ function participantsEqual(
         p.userId === b[i].userId &&
         p.speaking === b[i].speaking &&
         p.audioEnabled === b[i].audioEnabled &&
-        p.videoEnabled === b[i].videoEnabled,
+        p.videoEnabled === b[i].videoEnabled &&
+        p.speakingWhileMuted === b[i].speakingWhileMuted,
     )
   );
 }
@@ -66,34 +73,40 @@ export function startLotusCallState(vm: CallViewModel): () => void {
   // lotusDecorations.ts.
   if (!widget) return () => undefined;
 
-  const sub: Subscription = vm.userMedia$
-    .pipe(
-      switchMap((members) =>
-        members.length === 0
-          ? of([] as ParticipantState[])
-          : combineLatest(
-              members.map((m) =>
-                combineLatest([
-                  m.speaking$,
-                  m.audioEnabled$,
-                  m.videoEnabled$,
-                ]).pipe(
-                  map(
-                    ([
-                      speaking,
-                      audioEnabled,
-                      videoEnabled,
-                    ]): ParticipantState => ({
-                      id: m.id,
-                      userId: m.userId,
-                      speaking,
-                      audioEnabled,
-                      videoEnabled,
-                    }),
-                  ),
-                ),
+  const participants$ = vm.userMedia$.pipe(
+    switchMap((members) =>
+      members.length === 0
+        ? of([] as (ParticipantState & { local: boolean })[])
+        : combineLatest(
+            members.map((m) =>
+              combineLatest([
+                m.speaking$,
+                m.audioEnabled$,
+                m.videoEnabled$,
+              ]).pipe(
+                map(([speaking, audioEnabled, videoEnabled]) => ({
+                  id: m.id,
+                  userId: m.userId,
+                  speaking,
+                  audioEnabled,
+                  videoEnabled,
+                  local: m.local === true,
+                })),
               ),
             ),
+          ),
+    ),
+  );
+
+  const sub: Subscription = combineLatest([
+    participants$,
+    observeSpeakingWhileMuted$(vm).pipe(startWith(false)),
+  ])
+    .pipe(
+      map(([members, speakingWhileMuted]): ParticipantState[] =>
+        members.map(({ local, ...p }) =>
+          local ? { ...p, speakingWhileMuted } : p,
+        ),
       ),
       // `speaking` flips rapidly; drop no-op repeats BEFORE throttling so
       // the throttle window isn't spent re-emitting an unchanged value.

src/lotus/lotusMutedSpeech.test.ts

@@ -0,0 +1,27 @@
+/*
+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 { expect, test } from "vitest";
+
+import { MutedSpeechGate } from "./lotusMutedSpeech";
+
+test("needs 300 ms of voice to switch on and 800 ms of quiet to switch off", () => {
+  const g = new MutedSpeechGate(0.015);
+  expect(g.push(0.1)).toBe(false);
+  expect(g.push(0.1)).toBe(false);
+  expect(g.push(0.1)).toBe(true);
+  // a short dip does not drop it
+  for (let i = 0; i < 7; i++) expect(g.push(0.0)).toBe(true);
+  expect(g.push(0.0)).toBe(false);
+  // one loud sample after quiet does not re-trigger
+  expect(g.push(0.2)).toBe(false);
+});
+
+test("keyboard-level noise below the threshold never triggers", () => {
+  const g = new MutedSpeechGate(0.015);
+  for (let i = 0; i < 50; i++) expect(g.push(0.01)).toBe(false);
+});

src/lotus/lotusMutedSpeech.ts

@@ -0,0 +1,154 @@
+/*
+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 LocalTrackPublication,
+  type Room as LivekitRoom,
+  RoomEvent,
+  Track,
+} from "livekit-client";
+import { Observable, distinctUntilChanged, switchMap } from "rxjs";
+
+import { type CallViewModel } from "../state/CallViewModel/CallViewModel";
+
+/**
+ * [lotus #37] "Talking while muted" detection for the LOCAL participant.
+ *
+ * While the mic is muted LiveKit keeps the capture alive and merely disables
+ * the published MediaStreamTrack, so VAD/`speaking` is false and nothing
+ * tells the user they are talking into a muted mic. This taps a CLONE of the
+ * published track (the post-processor track when the in-source denoiser is
+ * active, so keyboard noise doesn't count), samples RMS at ~10 Hz and emits a
+ * debounced boolean. Zero cost when unmuted (tap torn down), local-only —
+ * the flag rides `io.lotus.call_state` to the host and never reaches other
+ * participants.
+ */
+
+export const MUTED_SPEECH_RMS = 0.015; // ≈ −36 dBFS; normal speech into a headset is 0.05–0.3
+const SAMPLE_MS = 100;
+const ON_SAMPLES = 3; // 300 ms of voice before we say "talking"
+const OFF_SAMPLES = 8; // 800 ms of quiet before we drop it
+
+/** Pure hysteresis gate over successive RMS samples (unit-tested). */
+export class MutedSpeechGate {
+  private above = 0;
+
+  private below = 0;
+
+  private on = false;
+
+  public constructor(private readonly threshold = MUTED_SPEECH_RMS) {}
+
+  public push(rms: number): boolean {
+    if (rms >= this.threshold) {
+      this.above += 1;
+      this.below = 0;
+      if (!this.on && this.above >= ON_SAMPLES) this.on = true;
+    } else {
+      this.below += 1;
+      this.above = 0;
+      if (this.on && this.below >= OFF_SAMPLES) this.on = false;
+    }
+    return this.on;
+  }
+
+  public get value(): boolean {
+    return this.on;
+  }
+}
+
+const mutedMicTrack = (room: LivekitRoom): MediaStreamTrack | null => {
+  const pub: LocalTrackPublication | undefined =
+    room.localParticipant.getTrackPublication(Track.Source.Microphone);
+  const track = pub?.track?.mediaStreamTrack;
+  return pub?.isMuted && track && track.readyState === "live" ? track : null;
+};
+
+/**
+ * Emits while the local mic is muted and voice is detected on it. Emits
+ * `false` whenever the mic is unmuted, unpublished or the connection changes.
+ */
+export function observeSpeakingWhileMuted$(
+  vm: CallViewModel,
+): Observable<boolean> {
+  return vm.allConnections$.pipe(
+    switchMap(
+      (data) =>
+        new Observable<boolean>((subscriber) => {
+          const rooms = data.getConnections().map((c) => c.livekitRoom);
+          let ctx: AudioContext | null = null;
+          let clone: MediaStreamTrack | null = null;
+          let timer: ReturnType<typeof setInterval> | undefined;
+          let tapped: MediaStreamTrack | null = null;
+
+          const stopTap = (): void => {
+            if (timer !== undefined) clearInterval(timer);
+            timer = undefined;
+            clone?.stop();
+            clone = null;
+            void ctx?.close().catch(() => undefined);
+            ctx = null;
+            tapped = null;
+            subscriber.next(false);
+          };
+
+          const startTap = (source: MediaStreamTrack): void => {
+            try {
+              clone = source.clone();
+              clone.enabled = true; // the source is disabled by the mute — the clone must not be
+              ctx = new AudioContext();
+              const analyser = ctx.createAnalyser();
+              analyser.fftSize = 1024;
+              ctx
+                .createMediaStreamSource(new MediaStream([clone]))
+                .connect(analyser);
+              const buf = new Float32Array(analyser.fftSize);
+              const gate = new MutedSpeechGate();
+              tapped = source;
+              timer = setInterval(() => {
+                analyser.getFloatTimeDomainData(buf);
+                let sum = 0;
+                for (let i = 0; i < buf.length; i += 1) sum += buf[i] * buf[i];
+                subscriber.next(gate.push(Math.sqrt(sum / buf.length)));
+              }, SAMPLE_MS);
+            } catch {
+              stopTap();
+            }
+          };
+
+          const reconcile = (): void => {
+            const track =
+              rooms.map(mutedMicTrack).find((t) => t !== null) ?? null;
+            if (track === tapped) return;
+            if (tapped) stopTap();
+            if (track) startTap(track);
+          };
+
+          const events = [
+            RoomEvent.TrackMuted,
+            RoomEvent.TrackUnmuted,
+            RoomEvent.LocalTrackPublished,
+            RoomEvent.LocalTrackUnpublished,
+            RoomEvent.Reconnected,
+            RoomEvent.Disconnected,
+          ] as const;
+          rooms.forEach((room) =>
+            events.forEach((ev) => room.on(ev, reconcile)),
+          );
+          subscriber.next(false);
+          reconcile();
+          return () => {
+            rooms.forEach((room) =>
+              events.forEach((ev) => room.off(ev, reconcile)),
+            );
+            if (tapped) stopTap();
+          };
+        }),
+    ),
+    distinctUntilChanged(),
+  );
+}