diff --git a/sdk/src/media-backend/livekit/LivekitMediaTrack.test.ts b/sdk/src/media-backend/livekit/LivekitMediaTrack.test.ts
index fc11f300d..790eddd1a 100644
--- a/sdk/src/media-backend/livekit/LivekitMediaTrack.test.ts
+++ b/sdk/src/media-backend/livekit/LivekitMediaTrack.test.ts
@@ -225,6 +225,51 @@ describe("createLivekitMediaTrack", () => {
     );
   });
 
+  // LiveKit's media stream track is the processed one while a processor runs,
+  // and that one names no device; the camera behind it does
+  it.each([
+    { deviceId: "cam-1", switches: false },
+    { deviceId: "cam-2", switches: true },
+  ])(
+    "leaves a processed camera alone unless its device changes: $deviceId",
+    async ({ deviceId, switches }) => {
+      vi.stubGlobal("MediaStreamTrack", class {});
+      onTestFinished(() => {
+        vi.unstubAllGlobals();
+      });
+      const camera = Object.assign(Object.create(LocalVideoTrack.prototype), {
+        source: Track.Source.Camera,
+        _mediaStreamTrack: {
+          readyState: "live",
+          label: "",
+          getSettings: () => ({ deviceId: "cam-1" }),
+        },
+        // What LiveKit shows as the track, with no device of its own
+        processor: {
+          processedTrack: { label: "", getSettings: () => ({}) },
+        },
+      }) as LocalVideoTrack;
+      const switchActiveDevice = vi.fn().mockResolvedValue(true);
+      const { participant, publication, room } = fakes({
+        track: camera as unknown as FakeTrack,
+      });
+      const track = createLocalLivekitMediaTrack(
+        scope,
+        Object.assign(participant, {
+          isLocal: true,
+        }) as unknown as LocalParticipant,
+        publication as unknown as TrackPublication,
+        Object.assign(room, { switchActiveDevice }),
+        vi.fn(),
+      );
+
+      await track.setDevice(deviceId);
+      expect(switchActiveDevice.mock.calls).toEqual(
+        switches ? [["videoinput", deviceId, false]] : [],
+      );
+    },
+  );
+
   // LiveKit stops a camera that is turned off; the processor is built when it
   // is turned on again
   it.each([
diff --git a/sdk/src/media-backend/livekit/LivekitMediaTrack.ts b/sdk/src/media-backend/livekit/LivekitMediaTrack.ts
index 588c55761..c32fc2d52 100644
--- a/sdk/src/media-backend/livekit/LivekitMediaTrack.ts
+++ b/sdk/src/media-backend/livekit/LivekitMediaTrack.ts
@@ -199,11 +199,13 @@ export function createLocalLivekitMediaTrack(
         throw new MatrixRTCError(`A ${base.source} track has no device`);
       // LiveKit decides whether to restart by the constraints it keeps, which
       // a browser that does not echo the device id back (Firefox) leaves
-      // without one; a track already on the device is left alone here
+      // without one; a track already on the device is left alone here. Asked
+      // of the camera track itself: with a processor on, LiveKit's media
+      // stream track is the processed one, which names no device
       const current = publication.track;
       if (
         current instanceof LocalTrack &&
-        current.mediaStreamTrack.getSettings().deviceId === deviceId
+        (await current.getDeviceId(false)) === deviceId
       )
         return;
       await room.switchActiveDevice(