Creating the silence based on the chunk size.

src/pipecat/services/tavus/video.py

@@ -202,11 +202,12 @@ class TavusVideoService(AIService):
         # limit (even including base64 encoding). For a sample rate of 16000,
         # that would be 32000 / 20 = 1600.
         sample_rate = self._client.out_sample_rate
+        # 50 ms of audio
         MAX_CHUNK_SIZE = int((sample_rate * 2) / 20)
 
         audio_buffer = bytearray()
         current_idx_str = None
-        silence = b"\x00\x00"
+        silence = b"\x00" * MAX_CHUNK_SIZE
         samples_sent = 0
         start_time = None
 

src/pipecat/transports/services/tavus.py

@@ -405,7 +405,8 @@ class TavusOutputTransport(BaseOutputTransport):
                 self._start_time = time.time()
                 self._samples_sent = 0
             elif isinstance(frame, BotStoppedSpeakingFrame):
-                await self._client.encode_audio_and_send(b"\x00\x00", True, self._current_idx_str)
+                silence = b"\x00" * self.audio_chunk_size
+                await self._client.encode_audio_and_send(silence, True, self._current_idx_str)
                 self._current_idx_str = None
         await super().push_frame(frame, direction)