diff --git a/gateway/run.py b/gateway/run.py
index 29d6c9dd4103..0b7a9ff4508b 100644
--- a/gateway/run.py
+++ b/gateway/run.py
@@ -21966,6 +21966,20 @@ def _should_send_voice_reply(
             )
             return False
 
+        # Desktop speaks for itself: the desktop app's useAutoSpeakReplies
+        # hook calls playSpeechText for the same reply, so a gateway voice
+        # message here makes every auto-TTS reply play twice (two independent
+        # synthesizations of the same text — logged as one "Generating speech"
+        # plus two "TTS audio saved" entries). The desktop path is the
+        # authoritative one for that surface; the gateway has no way to know
+        # whether the client spoke, so it must stay out of the way (#90297).
+        if event.source.platform.value == "desktop":
+            logger.debug(
+                "Auto voice reply skipped: desktop surface speaks locally (mode=%s chat=%s)",
+                voice_mode, chat_id,
+            )
+            return False
+
         # Dedup: agent already called TTS tool in THIS turn only
         last_user_idx = None
         for i, msg in enumerate(reversed(agent_messages)):
diff --git a/tests/gateway/test_desktop_no_gateway_voice_reply.py b/tests/gateway/test_desktop_no_gateway_voice_reply.py
new file mode 100644
index 000000000000..da8f09d151f9
--- /dev/null
+++ b/tests/gateway/test_desktop_no_gateway_voice_reply.py
@@ -0,0 +1,101 @@
+"""The gateway must not send an auto-TTS voice reply to the desktop surface (#90297).
+
+When `voice.auto_tts` is on, two independent TTS paths fire for the same
+reply on the desktop: the gateway's `_send_voice_reply` (via
+`adapter.send_voice`) and the desktop app's `useAutoSpeakReplies` hook (via
+`/api/audio/speak` → `playSpeechText`). Neither knows about the other, so
+every reply plays twice (one "Generating speech", two "TTS audio saved").
+The desktop hook is the authoritative speaker for that surface; the gateway
+side now steps out.
+
+The desktop platform value is produced by the session routing layer at
+runtime (it is not a static ``Platform`` enum member), so the tests stub a
+platform object whose ``.value`` is ``"desktop"`` — exactly the comparison
+``_should_send_voice_reply`` performs.
+"""
+
+from unittest.mock import MagicMock
+
+from gateway.config import Platform
+from gateway.platforms.base import MessageEvent, MessageType
+from gateway.run import GatewayRunner
+from gateway.session import SessionSource
+
+
+class _DesktopPlatform(str):
+    """Hashable stand-in for the runtime desktop platform value.
+
+    The desktop platform is produced by the session routing layer at
+    runtime (not a static ``Platform`` enum member), so the tests stub a
+    hashable object whose ``.value`` is ``"desktop"`` — exactly the
+    comparison and dict-key usage ``_should_send_voice_reply`` performs.
+    """
+
+    @property
+    def value(self) -> str:
+        return self
+
+
+def _desktop_platform() -> _DesktopPlatform:
+    return _DesktopPlatform("desktop")
+
+
+def _make_runner() -> GatewayRunner:
+    runner = GatewayRunner.__new__(GatewayRunner)
+    runner.adapters = {}
+    runner._voice_mode = {}
+    return runner
+
+
+def _make_event(platform) -> MessageEvent:
+    return MessageEvent(
+        text="trigger",
+        source=SessionSource(
+            platform=platform,
+            chat_id="123",
+            user_id="u1",
+            user_name="User",
+        ),
+        message_type=MessageType.TEXT,
+        message_id="456",
+    )
+
+
+class TestDesktopNeverGetsGatewayVoiceReply:
+    def test_desktop_platform_skipped_even_with_auto_tts_on(self):
+        """The #90297 shape: global auto_tts on and the desktop would
+        otherwise qualify — the gateway must still stay silent because the
+        desktop's useAutoSpeakReplies speaks the same reply itself."""
+        runner = _make_runner()
+        adapter = MagicMock()
+        adapter._should_auto_tts_for_chat = MagicMock(return_value=True)
+        runner.adapters[_desktop_platform()] = adapter
+        runner._voice_mode = {}  # no explicit mode: adapter_auto_tts governs
+
+        event = _make_event(_desktop_platform())
+        assert (
+            runner._should_send_voice_reply(event, "spoken twice?", [])
+            is False
+        )
+
+    def test_desktop_skipped_with_explicit_all_mode(self):
+        """Even an explicit /voice all for the chat must not double-speak on
+        the desktop — the hook fires on every reply regardless of mode."""
+        runner = _make_runner()
+        runner.adapters["desktop"] = MagicMock()
+        runner._voice_mode = {("desktop", "123"): "all"}
+
+        event = _make_event(_desktop_platform())
+        assert runner._should_send_voice_reply(event, "hello", []) is False
+
+    def test_other_platforms_still_send(self):
+        """Guard: the skip is desktop-only — Telegram with auto_tts on still
+        gets its gateway voice reply."""
+        runner = _make_runner()
+        adapter = MagicMock()
+        adapter._should_auto_tts_for_chat = MagicMock(return_value=True)
+        runner.adapters[Platform.TELEGRAM] = adapter
+        runner._voice_mode = {}
+
+        event = _make_event(Platform.TELEGRAM)
+        assert runner._should_send_voice_reply(event, "hello", []) is True