feat(gladia): add VAD events support

Add support for Gladia's speech_start/speech_end events to emit
UserStartedSpeakingFrame and UserStoppedSpeakingFrame frames.

When enable_vad=True in GladiaInputParams:
- speech_start triggers interruption and pushes UserStartedSpeakingFrame
- speech_end pushes UserStoppedSpeakingFrame
- Tracks speaking state to prevent duplicate events

This allows using Gladia's built-in VAD instead of a separate VAD
in the pipeline.

src/pipecat/services/gladia/config.py

@@ -169,6 +169,9 @@ class GladiaInputParams(BaseModel):
         pre_processing: Audio pre-processing options
         realtime_processing: Real-time processing features
         messages_config: WebSocket message filtering options
+        enable_vad: Enable VAD to trigger end of utterance detection. This should be used
+            without any other VAD enabled in the agent and will emit the speaker started
+            and stopped frames. Defaults to False.
     """
 
     encoding: Optional[str] = "wav/pcm"
@@ -182,3 +185,4 @@ class GladiaInputParams(BaseModel):
     pre_processing: Optional[PreProcessingConfig] = None
     realtime_processing: Optional[RealtimeProcessingConfig] = None
     messages_config: Optional[MessagesConfig] = None
+    enable_vad: bool = False

src/pipecat/services/gladia/stt.py

@@ -28,6 +28,8 @@ from pipecat.frames.frames import (
     StartFrame,
     TranscriptionFrame,
     TranslationFrame,
+    UserStartedSpeakingFrame,
+    UserStoppedSpeakingFrame,
 )
 from pipecat.services.gladia.config import GladiaInputParams
 from pipecat.services.stt_service import WebsocketSTTService
@@ -266,6 +268,9 @@ class GladiaSTTService(WebsocketSTTService):
         self._max_buffer_size = max_buffer_size
         self._buffer_lock = asyncio.Lock()
 
+        # VAD state tracking
+        self._is_speaking = False
+
     def __str__(self):
         return f"{self.name} [{self._session_id}]"
 
@@ -507,6 +512,30 @@ class GladiaSTTService(WebsocketSTTService):
         await self.stop_ttfb_metrics()
         await self.stop_processing_metrics()
 
+    async def _on_speech_started(self):
+        """Handle speech start event from Gladia.
+
+        Triggers interruption and emits UserStartedSpeakingFrame when VAD is enabled.
+        """
+        if not self._params.enable_vad or self._is_speaking:
+            return
+        logger.debug(f"{self} User started speaking")
+        self._is_speaking = True
+        # Push interruption first to stop the bot, then notify about user speaking
+        await self.push_interruption_task_frame_and_wait()
+        await self.push_frame(UserStartedSpeakingFrame())
+
+    async def _on_speech_ended(self):
+        """Handle speech end event from Gladia.
+
+        Emits UserStoppedSpeakingFrame when VAD is enabled.
+        """
+        if not self._params.enable_vad or not self._is_speaking:
+            return
+        self._is_speaking = False
+        await self.push_frame(UserStoppedSpeakingFrame())
+        logger.debug(f"{self} User stopped speaking")
+
     async def _send_audio(self, audio: bytes):
         """Send audio chunk with proper message format."""
         if self._websocket and self._websocket.state is State.OPEN:
@@ -599,6 +628,10 @@ class GladiaSTTService(WebsocketSTTService):
                                 translation, "", time_now_iso8601(), translated_language
                             )
                         )
+                elif content["type"] == "speech_start":
+                    await self._on_speech_started()
+                elif content["type"] == "speech_end":
+                    await self._on_speech_ended()
             except json.JSONDecodeError:
                 logger.warning(f"{self} Received non-JSON message: {message}")