diff --git a/src/pipecat/services/gemini_multimodal_live/gemini.py b/src/pipecat/services/gemini_multimodal_live/gemini.py
index 57753e9cd..0241b3b64 100644
--- a/src/pipecat/services/gemini_multimodal_live/gemini.py
+++ b/src/pipecat/services/gemini_multimodal_live/gemini.py
@@ -288,6 +288,10 @@ class GeminiMultimodalLiveLLMService(LLMService):
         )
 
     async def _handle_transcribe_model_audio(self, audio, context):
+        # Early return if modalities are not set to audio.
+        if self._settings["modalities"] != GeminiMultimodalModalities.AUDIO:
+            return
+
         text = await self._transcribe_audio(audio, context)
         logger.debug(f"[Transcription:model] {text}")
         # We add user messages directly to the context. We don't do that for assistant messages,