LLMConfigureOutputFrame: allow configuring LLM output

src/pipecat/frames/frames.py

@@ -607,6 +607,21 @@ class LLMEnablePromptCachingFrame(DataFrame):
     enable: bool
 
 
+@dataclass
+class LLMConfigureOutputFrame(DataFrame):
+    """Frame to configure LLM output.
+
+    This frame is used to configure how the LLM produces output. For example, it
+    can tell the LLM to generate tokens that should be added to the context but
+    not spoken by the TTS service (if one is present in the pipeline).
+
+    Parameters:
+        skip_tts: Whether LLM tokens should skip the TTS service (if any).
+    """
+
+    skip_tts: bool
+
+
 @dataclass
 class TTSSpeakFrame(DataFrame):
     """Frame containing text that should be spoken by TTS.

src/pipecat/processors/aggregators/dtmf_aggregator.py

@@ -103,7 +103,7 @@ class DTMFAggregator(FrameProcessor):
         digit_value = frame.button.value
         self._aggregation += digit_value
 
-        # For first digit, schedule interruption in separate task
+        # For first digit, schedule interruption.
         if is_first_digit:
             await self.push_frame(BotInterruptionFrame(), FrameDirection.UPSTREAM)
 

src/pipecat/services/llm_service.py

@@ -37,6 +37,8 @@ from pipecat.frames.frames import (
     FunctionCallResultFrame,
     FunctionCallResultProperties,
     FunctionCallsStartedFrame,
+    LLMConfigureOutputFrame,
+    LLMTextFrame,
     StartFrame,
     StartInterruptionFrame,
     UserImageRequestFrame,
@@ -179,6 +181,7 @@ class LLMService(AIService):
         self._function_call_tasks: Dict[asyncio.Task, FunctionCallRunnerItem] = {}
         self._sequential_runner_task: Optional[asyncio.Task] = None
         self._tracing_enabled: bool = False
+        self._skip_tts: bool = False
 
         self._register_event_handler("on_function_calls_started")
         self._register_event_handler("on_completion_timeout")
@@ -272,6 +275,20 @@ class LLMService(AIService):
 
         if isinstance(frame, StartInterruptionFrame):
             await self._handle_interruptions(frame)
+        elif isinstance(frame, LLMConfigureOutputFrame):
+            self._skip_tts = frame.skip_tts
+
+    async def push_frame(self, frame: Frame, direction: FrameDirection = FrameDirection.DOWNSTREAM):
+        """Pushes a frame.
+
+        Args:
+            frame: The frame to push.
+            direction: The direction of frame pushing.
+        """
+        if isinstance(frame, LLMTextFrame):
+            frame.skip_tts = self._skip_tts
+
+        await super().push_frame(frame, direction)
 
     async def _handle_interruptions(self, _: StartInterruptionFrame):
         for function_name, entry in self._functions.items():