# Navigate to Twilio examplecd examples/realtime/twilio# Install dependenciesuv sync# Start the serveruv run server.py# In another terminal, start ngrokngrok http 5050# Update Twilio webhook to ngrok URL# Call the Twilio number

# Lines 181-194: Buffered audio handling ✅self._incoming_audio_buffer.append(audio_data)asyncdef_buffer_flush_loop(self):whileTrue:awaitasyncio.sleep(0.1)ifself._incoming_audio_buffer:# Flush accumulated audio to OpenAIself._flush_incoming_audio()

# Lines 152-158: NO BUFFERING! ❌ifevent.type=="audio_chunk":audio_data=base64.b64encode(event.audio).decode()awaitself.send_twilio_message({"event":"media","media": {"payload":audio_data}# Sent immediately!    })

ifevent.type=="audio_chunk":print(f"Chunk size:{len(event.audio)} bytes")# Typical output:# Chunk size: 20 bytes  ← TOO SMALL!# Chunk size: 40 bytes  ← TOO SMALL!# Chunk size: 60 bytes  ← TOO SMALL!# ...

classTwilioRealtimeHandler:def__init__(self, ...):# Existing incoming bufferself._incoming_audio_buffer:list[bytes]= []# NEW: Add outgoing bufferself._outgoing_audio_buffer:list[bytes]= []# ✅ Added this# Track buffered marks for proper cleanupself._buffered_marks:set[str]=set()# ✅ Added this

ifevent.type=="audio_chunk":# Send immediately - causes jittering! ❌audio_data=base64.b64encode(event.audio).decode()awaitself.send_twilio_message({"event":"media","media": {"payload":audio_data}    })

ifevent.type=="audio_chunk":# Buffer the audio chunk ✅self._outgoing_audio_buffer.append(event.audio)# Flush if buffer is large enough (50ms worth of data)# At 8kHz with g711_ulaw, 50ms = 400 bytestotal_size=sum(len(chunk)forchunkinself._outgoing_audio_buffer)iftotal_size>=400:awaitself._flush_outgoing_audio_buffer()

asyncdef_flush_outgoing_audio_buffer(self):"""Flush accumulated outgoing audio to Twilio"""ifnotself._outgoing_audio_buffer:return# Combine all buffered chunkscombined_audio=b"".join(self._outgoing_audio_buffer)# Clear the bufferself._outgoing_audio_buffer.clear()# Encode and send to Twilioaudio_data=base64.b64encode(combined_audio).decode()awaitself.send_twilio_message({"event":"media","media": {"payload":audio_data}    })# Send all buffered marksformark_idinself._buffered_marks:awaitself.send_twilio_message({"event":"mark","mark": {"name":mark_id}        })self._buffered_marks.clear()

asyncdef_buffer_flush_loop(self):"""Periodically flush both incoming and outgoing audio buffers"""whileTrue:awaitasyncio.sleep(0.1)# Every 100ms# Flush incoming audio (Twilio → OpenAI)ifself._incoming_audio_buffer:awaitself._flush_incoming_audio()# Flush outgoing audio (OpenAI → Twilio) ✅ NEWifself._outgoing_audio_buffer:awaitself._flush_outgoing_audio_buffer()

elifevent.type=="audio_end":# Flush any remaining outgoing audio ✅ifself._outgoing_audio_buffer:awaitself._flush_outgoing_audio_buffer()awaitself.send_twilio_message({"event":"clear"})elifevent.type=="audio_interrupted":# Flush before clearing ✅ifself._outgoing_audio_buffer:awaitself._flush_outgoing_audio_buffer()awaitself.send_twilio_message({"event":"clear"})

elifevent.type=="audio_transcript_done":# Buffer the mark instead of sending immediatelymark_id=event.item_idself._buffered_marks.add(mark_id)# ✅ Track for later sending

# Restart the server with the fixuv run server.py# Call the Twilio number again# Listen to the audio quality

asyncdef_flush_outgoing_audio_buffer(self):ifnotself._outgoing_audio_buffer:returncombined_audio=b"".join(self._outgoing_audio_buffer)print(f"Sending buffered audio:{len(combined_audio)} bytes")# Log# Output:# Sending buffered audio: 480 bytes  ✅ Good size!# Sending buffered audio: 520 bytes  ✅ Good size!# Sending buffered audio: 440 bytes  ✅ Good size!

importasyncioclassTestBuffer:def__init__(self):self._outgoing_audio_buffer= []self._buffered_marks=set()asyncdefadd_chunk(self,data:bytes):"""Simulate receiving audio chunk from OpenAI"""self._outgoing_audio_buffer.append(data)total_size=sum(len(chunk)forchunkinself._outgoing_audio_buffer)print(f"Buffer size:{total_size} bytes")iftotal_size>=400:awaitself.flush()asyncdefflush(self):"""Flush buffered audio"""ifnotself._outgoing_audio_buffer:returncombined=b"".join(self._outgoing_audio_buffer)print(f"✅ Flushing{len(combined)} bytes")self._outgoing_audio_buffer.clear()asyncdefmain():buffer=TestBuffer()print("[Test 1] Small chunks accumulate before flushing")awaitbuffer.add_chunk(b"X"*50)# 50 bytesawaitbuffer.add_chunk(b"X"*50)# 100 bytes totalawaitbuffer.add_chunk(b"X"*50)# 150 bytes totalawaitbuffer.add_chunk(b"X"*50)# 200 bytes totalawaitbuffer.add_chunk(b"X"*50)# 250 bytes totalawaitbuffer.add_chunk(b"X"*50)# 300 bytes totalawaitbuffer.add_chunk(b"X"*50)# 350 bytes totalawaitbuffer.add_chunk(b"X"*100)# 450 bytes → FLUSH! ✅print("\n[Test 2] Large chunk triggers immediate flush")awaitbuffer.add_chunk(b"X"*500)# 500 bytes → FLUSH! ✅print("\n[Test 3] Multiple small then flush")awaitbuffer.add_chunk(b"X"*100)# 100 bytesawaitbuffer.add_chunk(b"X"*100)# 200 bytesawaitbuffer.flush()# Manual flush ✅asyncio.run(main())

[Test 1] Small chunks accumulate before flushingBuffer size: 50 bytesBuffer size: 100 bytesBuffer size: 150 bytesBuffer size: 200 bytesBuffer size: 250 bytesBuffer size: 300 bytesBuffer size: 350 bytesBuffer size: 450 bytes✅ Flushing 450 bytes[Test 2] Large chunk triggers immediate flushBuffer size: 500 bytes✅ Flushing 500 bytes[Test 3] Multiple small then flushBuffer size: 100 bytesBuffer size: 200 bytes✅ Flushing 200 bytes

# Lintinguv run ruff check examples/realtime/twilio/twilio_handler.py# Type checkinguv run mypy examples/realtime/twilio/twilio_handler.py# Formattinguv run ruff format examples/realtime/twilio/twilio_handler.py

✅ Linting: No issues✅ Type checking: No errors✅ Formatting: All files formatted