diff --git a/src/lotus/lotusDenoise.ts b/src/lotus/lotusDenoise.ts
new file mode 100644
index 00000000..623c986a
--- /dev/null
+++ b/src/lotus/lotusDenoise.ts
@@ -0,0 +1,98 @@
+/*
+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.
+ */
+export function startLotusDenoise(vm: CallViewModel): () => void {
+  if (lotusParam("lotusDenoise") !== "ml") return () => undefined;
+
+  const config: LotusDenoiseConfig = {
+    model: lotusParam("lotusModel") === "speex" ? "speex" : "rnnoise",
+    assetBase: lotusParam("lotusDenoiseBase") || "./denoise/",
+    gate: lotusFlag("lotusGate"),
+    gateThreshold: Number(lotusParam("lotusGateThreshold")) || -50,
+  };
+
+  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();
+    }
+  };
+}
diff --git a/src/lotus/lotusDenoiseProcessor.ts b/src/lotus/lotusDenoiseProcessor.ts
new file mode 100644
index 00000000..f39a3273
--- /dev/null
+++ b/src/lotus/lotusDenoiseProcessor.ts
@@ -0,0 +1,151 @@
+/*
+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 AudioProcessorOptions,
+  type Track,
+  type TrackProcessor,
+} from "livekit-client";
+import { logger } from "matrix-js-sdk/lib/logger";
+
+export type LotusDenoiseModel = "rnnoise" | "speex";
+
+export interface LotusDenoiseConfig {
+  model: LotusDenoiseModel;
+  /** Base URL the worklet scripts + wasm are served from (e.g. "./denoise/"). */
+  assetBase: string;
+  gate: boolean;
+  gateThreshold: number;
+}
+
+// Flat sapphi worklets: each registers a processor under these names when its
+// script module is added. Same assets the Lotus host already ships under
+// public/element-call/denoise/, so no new EC dependency is needed.
+const PROCESSORS: Record<
+  LotusDenoiseModel,
+  { name: string; script: string; wasm: string }
+> = {
+  rnnoise: {
+    name: "@sapphi-red/web-noise-suppressor/rnnoise",
+    script: "rnnoiseWorklet.js",
+    wasm: "rnnoise.wasm",
+  },
+  speex: {
+    name: "@sapphi-red/web-noise-suppressor/speex",
+    script: "speexWorklet.js",
+    wasm: "speex.wasm",
+  },
+};
+const GATE_NAME = "@sapphi-red/web-noise-suppressor/noiseGate";
+const GATE_SCRIPT = "noiseGateWorklet.js";
+
+/**
+ * A LiveKit audio TrackProcessor that runs Lotus ML noise suppression
+ * (RNNoise/Speex) on the local microphone track, as a first-class stage in
+ * Element Call's publish pipeline.
+ *
+ * This replaces the host's `getUserMedia` monkeypatch: because it's a real
+ * LiveKit processor, EC re-applies it on every (re)publish, so denoise
+ * survives EC's own mid-call reconnect — the root cause of the A7
+ * "mic dead after reconnect" bug.
+ */
+export class LotusDenoiseProcessor
+  implements TrackProcessor<Track.Kind.Audio, AudioProcessorOptions>
+{
+  public readonly name = "lotus-denoise";
+  public processedTrack?: MediaStreamTrack;
+
+  private source?: MediaStreamAudioSourceNode;
+  private nodes: AudioWorkletNode[] = [];
+
+  public constructor(private readonly config: LotusDenoiseConfig) {}
+
+  public async init(opts: AudioProcessorOptions): Promise<void> {
+    await this.build(opts.audioContext, opts.track);
+  }
+
+  public async restart(opts: AudioProcessorOptions): Promise<void> {
+    this.teardownGraph();
+    await this.build(opts.audioContext, opts.track);
+  }
+
+  public async destroy(): Promise<void> {
+    this.teardownGraph();
+  }
+
+  private async build(
+    ctx: AudioContext,
+    track: MediaStreamTrack,
+  ): Promise<void> {
+    const base = this.config.assetBase;
+    const proc = PROCESSORS[this.config.model];
+
+    // Register worklet modules (idempotent) and load the wasm binary.
+    await ctx.audioWorklet.addModule(base + proc.script);
+    if (this.config.gate) await ctx.audioWorklet.addModule(base + GATE_SCRIPT);
+    const wasmBinary = await fetch(base + proc.wasm).then((r) => {
+      if (!r.ok) throw new Error(`denoise wasm ${r.status}`);
+      return r.arrayBuffer();
+    });
+
+    const source = ctx.createMediaStreamSource(new MediaStream([track]));
+    const dest = ctx.createMediaStreamDestination();
+    let head: AudioNode = source;
+
+    if (this.config.gate) {
+      const gate = new AudioWorkletNode(ctx, GATE_NAME, {
+        processorOptions: {
+          openThreshold: this.config.gateThreshold,
+          closeThreshold: this.config.gateThreshold - 5,
+          holdMs: 150,
+          maxChannels: 1,
+        },
+      });
+      head.connect(gate);
+      head = gate;
+      this.nodes.push(gate);
+    }
+
+    const ml = new AudioWorkletNode(ctx, proc.name, {
+      channelCount: 1,
+      numberOfInputs: 1,
+      numberOfOutputs: 1,
+      processorOptions: { maxChannels: 1, wasmBinary },
+    });
+    head.connect(ml);
+    ml.connect(dest);
+    this.nodes.push(ml);
+
+    this.source = source;
+    this.processedTrack = dest.stream.getAudioTracks()[0];
+    logger.info(`[lotus] denoise processor active (${this.config.model})`);
+  }
+
+  private teardownGraph(): void {
+    for (const node of this.nodes) {
+      try {
+        node.port.postMessage("destroy");
+      } catch {
+        /* ignore */
+      }
+      try {
+        node.disconnect();
+      } catch {
+        /* ignore */
+      }
+    }
+    this.nodes = [];
+    try {
+      this.source?.disconnect();
+    } catch {
+      /* ignore */
+    }
+    this.source = undefined;
+    this.processedTrack?.stop();
+    this.processedTrack = undefined;
+  }
+}
diff --git a/src/room/InCallView.tsx b/src/room/InCallView.tsx
index fd1078b9..50c19f44 100644
--- a/src/room/InCallView.tsx
+++ b/src/room/InCallView.tsx
@@ -34,6 +34,7 @@ import { startLotusFocus } from "../lotus/lotusFocus";
 import { startLotusAudioInject } from "../lotus/lotusAudioInject";
 import { startLotusQuality } from "../lotus/lotusQuality";
 import { startLotusDecorations } from "../lotus/lotusDecorations";
+import { startLotusDenoise } from "../lotus/lotusDenoise";
 import styles from "./InCallView.module.css";
 import { GridTile } from "../tile/GridTile";
 import { SettingsModal, defaultSettingsTab } from "../settings/SettingsModal";
@@ -299,6 +300,9 @@ export const InCallView: FC<InCallViewProps> = ({
   // [lotus] Receive per-user avatar-decoration URLs from the host and render
   // them on in-call tile avatars (#6). No-op unless the host sends them.
   useEffect(() => startLotusDecorations(), []);
+  // [lotus] Apply ML denoise to the mic as a first-class audio processor that
+  // survives reconnects (#1 / A7). No-op unless lotusDenoise=ml.
+  useEffect(() => startLotusDenoise(vm), [vm]);
 
   const fatalCallError = useBehavior(vm.fatalError$);
   // Stop the rendering and throw for the error boundary