diff --git a/apps/desktop/src/lib/voice-playback.test.ts b/apps/desktop/src/lib/voice-playback.test.ts
new file mode 100644
index 000000000000..b3efa3676d37
--- /dev/null
+++ b/apps/desktop/src/lib/voice-playback.test.ts
@@ -0,0 +1,67 @@
+import { afterEach, describe, expect, it, vi } from 'vitest'
+
+import { startSpeechStream } from './voice-playback'
+
+class FakeWebSocket {
+  static CONNECTING = 0
+  static OPEN = 1
+  static CLOSED = 3
+  static instances: FakeWebSocket[] = []
+
+  binaryType: BinaryType = 'blob'
+  close = vi.fn(() => {
+    this.readyState = FakeWebSocket.CLOSED
+  })
+  onclose: ((event: CloseEvent) => void) | null = null
+  onerror: ((event: Event) => void) | null = null
+  onmessage: ((event: MessageEvent) => void) | null = null
+  onopen: ((event: Event) => void) | null = null
+  readyState = FakeWebSocket.CONNECTING
+  send = vi.fn()
+
+  constructor(readonly url: string) {
+    FakeWebSocket.instances.push(this)
+  }
+}
+
+describe('voice playback streaming', () => {
+  afterEach(() => {
+    FakeWebSocket.instances = []
+    vi.unstubAllGlobals()
+    delete (window as { hermesDesktop?: unknown }).hermesDesktop
+  })
+
+  it('falls back to plain audio playback when WebAudio cannot initialize', async () => {
+    const warn = vi.spyOn(console, 'warn').mockImplementation(() => undefined)
+    vi.stubGlobal('WebSocket', FakeWebSocket)
+    vi.stubGlobal(
+      'AudioContext',
+      class {
+        constructor() {
+          throw new Error('audio device unavailable')
+        }
+      }
+    )
+    ;(window as { hermesDesktop?: unknown }).hermesDesktop = {
+      getConnection: vi.fn().mockResolvedValue({ authMode: 'token', wsUrl: 'ws://local/api/ws?token=abc' })
+    }
+
+    const session = await startSpeechStream({ source: 'read-aloud' })
+
+    expect(session).not.toBeNull()
+    const socket = FakeWebSocket.instances[0]
+
+    expect(() => {
+      socket.onmessage?.(
+        new MessageEvent('message', {
+          data: JSON.stringify({ channels: 1, sample_rate: 24_000, type: 'start' })
+        })
+      )
+    }).not.toThrow()
+    await expect(session?.done).resolves.toBe('fallback')
+    expect(warn).toHaveBeenCalledWith(
+      'Voice playback streaming disabled: AudioContext unavailable',
+      expect.any(Error)
+    )
+  })
+})
diff --git a/apps/desktop/src/lib/voice-playback.ts b/apps/desktop/src/lib/voice-playback.ts
index 48a51d370817..ddc7b958c170 100644
--- a/apps/desktop/src/lib/voice-playback.ts
+++ b/apps/desktop/src/lib/voice-playback.ts
@@ -41,12 +41,25 @@ async function unlockAutoplay(): Promise<void> {
     return
   }
 
-  if (!unlockCtx) {
-    unlockCtx = new Ctor()
+  try {
+    if (!unlockCtx) {
+      unlockCtx = new Ctor()
+    }
+
+    if (unlockCtx.state === 'suspended') {
+      await unlockCtx.resume()
+    }
+  } catch (error) {
+    console.warn('Voice playback autoplay unlock skipped: AudioContext unavailable', error)
   }
+}
 
-  if (unlockCtx.state === 'suspended') {
-    await unlockCtx.resume()
+function createAudioContextOrNull(): AudioContext | null {
+  try {
+    return new AudioContext()
+  } catch (error) {
+    console.warn('Voice playback streaming disabled: AudioContext unavailable', error)
+    return null
   }
 }
 
@@ -272,7 +285,12 @@ function openSpeechStream(wsUrl: string, options: VoicePlaybackOptions): SpeechS
 
     if (frame.type === 'start') {
       streamRate = frame.sample_rate || 24_000
-      context = new AudioContext()
+      context = createAudioContextOrNull()
+
+      if (!context) {
+        settle(started ? 'done' : 'fallback')
+        return
+      }
 
       // Autoplay policy can hand back a suspended context when playback wasn't
       // started by a user gesture (e.g. a wake-word-started voice turn). Resume
diff --git a/hermes_cli/web_server.py b/hermes_cli/web_server.py
index a9a922aa5474..c16f02202fbd 100644
--- a/hermes_cli/web_server.py
+++ b/hermes_cli/web_server.py
@@ -4381,6 +4381,35 @@ def _fetch() -> Dict[str, Any]:
     return {"available": True, "voices": voices}
 
 
+DESKTOP_TTS_TIMEOUT_SECONDS = 75.0
+
+
+def _unlink_desktop_tts_path(file_path: Any) -> None:
+    if not file_path:
+        return
+    try:
+        Path(file_path).unlink(missing_ok=True)
+    except OSError:
+        pass
+
+
+def _cleanup_timed_out_desktop_tts(future: "asyncio.Future", output_path: Path) -> None:
+    """Remove audio files produced after the desktop speak request timed out."""
+    _unlink_desktop_tts_path(output_path)
+    try:
+        result_json = future.result()
+    except Exception:
+        return
+    try:
+        result = json.loads(result_json) if isinstance(result_json, str) else result_json
+    except Exception:
+        return
+    if isinstance(result, dict):
+        returned_path = result.get("file_path")
+        if returned_path and Path(returned_path) != output_path:
+            _unlink_desktop_tts_path(returned_path)
+
+
 @app.post("/api/audio/speak")
 async def speak_text(payload: TTSSpeakRequest, profile: Optional[str] = None):
     """Synthesize speech and return audio as base64 data URL.
@@ -4394,6 +4423,8 @@ async def speak_text(payload: TTSSpeakRequest, profile: Optional[str] = None):
     if not text:
         raise HTTPException(status_code=400, detail="Text is required")
 
+    output_path = Path(get_hermes_home()) / "audio_cache" / f"desktop_speak_{secrets.token_hex(8)}.mp3"
+
     try:
         from tools.tts_tool import text_to_speech_tool
 
@@ -4403,14 +4434,28 @@ def _speak_scoped():
             # resolution, so the task-local override inside this worker
             # thread is sufficient (same reasoning as the MCP probe scope).
             with _config_profile_scope(profile):
-                return text_to_speech_tool(text)
+                return text_to_speech_tool(text, output_path=str(output_path))
 
         loop = asyncio.get_running_loop()
-        result_json = await loop.run_in_executor(None, _speak_scoped)
+        speak_future = loop.run_in_executor(None, _speak_scoped)
+        result_json = await asyncio.wait_for(
+            asyncio.shield(speak_future),
+            timeout=DESKTOP_TTS_TIMEOUT_SECONDS,
+        )
     except HTTPException:
         # _config_profile_scope raises 400/404 for a bad profile — pass it
         # through instead of masking it as a 500 synthesis failure.
         raise
+    except (asyncio.TimeoutError, TimeoutError) as exc:
+        _log.warning(
+            "Desktop voice TTS timed out after %.1fs (text_len=%d)",
+            DESKTOP_TTS_TIMEOUT_SECONDS,
+            len(text),
+        )
+        speak_future.add_done_callback(
+            lambda future: _cleanup_timed_out_desktop_tts(future, output_path)
+        )
+        raise HTTPException(status_code=504, detail="Speech synthesis timed out") from exc
     except Exception as exc:
         _log.exception("Desktop voice TTS failed")
         raise HTTPException(status_code=500, detail=f"Speech synthesis failed: {exc}")
diff --git a/tests/hermes_cli/test_web_server_profile_unification.py b/tests/hermes_cli/test_web_server_profile_unification.py
index a0534ac27339..0b0d87a56839 100644
--- a/tests/hermes_cli/test_web_server_profile_unification.py
+++ b/tests/hermes_cli/test_web_server_profile_unification.py
@@ -444,6 +444,47 @@ def _fake_transcribe(path):
         assert resp.json()["transcript"] == "hi"
         assert seen["home"] == str(isolated_profiles["worker_beta"])
 
+    def test_speak_text_times_out_stalled_provider(self, client, monkeypatch, tmp_path):
+        import json
+        import threading
+        import time
+        from pathlib import Path
+
+        import hermes_cli.web_server as web_server
+        import tools.tts_tool as tts_tool
+
+        captured = {}
+        finished = threading.Event()
+
+        def stalled_tts(text, output_path=None):
+            time.sleep(0.05)
+            audio_file = Path(output_path)
+            captured["audio_file"] = audio_file
+            audio_file.parent.mkdir(parents=True, exist_ok=True)
+            audio_file.write_bytes(b"ID3late-audio-bytes")
+            finished.set()
+            return json.dumps({
+                "success": True,
+                "file_path": str(audio_file),
+                "provider": "test",
+            })
+
+        monkeypatch.setattr(tts_tool, "text_to_speech_tool", stalled_tts)
+        monkeypatch.setattr(web_server, "DESKTOP_TTS_TIMEOUT_SECONDS", 0.01)
+
+        resp = client.post("/api/audio/speak", json={"text": "hello there"})
+
+        assert resp.status_code == 504
+        assert "timed out" in resp.json()["detail"].lower()
+        assert finished.wait(timeout=1.0)
+
+        audio_file = captured["audio_file"]
+        for _ in range(20):
+            if not audio_file.exists():
+                break
+            time.sleep(0.01)
+        assert not audio_file.exists()
+
     def test_audio_endpoints_unknown_profile_404(self, client, isolated_profiles):
         resp = client.get("/api/audio/elevenlabs/voices?profile=ghost")
         assert resp.status_code == 404