|
| 1 | +import asyncio |
| 2 | +import queue |
| 3 | +import sys |
| 4 | +import threading |
| 5 | +from typing import Any |
| 6 | + |
| 7 | +import numpy as np |
| 8 | +import sounddevice as sd |
| 9 | + |
| 10 | +from agents import function_tool |
| 11 | +from agents.realtime import RealtimeAgent, RealtimeRunner, RealtimeSession, RealtimeSessionEvent |
| 12 | + |
| 13 | +# Audio configuration |
| 14 | +CHUNK_LENGTH_S = 0.05 # 50ms |
| 15 | +SAMPLE_RATE = 24000 |
| 16 | +FORMAT = np.int16 |
| 17 | +CHANNELS = 1 |
| 18 | + |
| 19 | +# Set up logging for OpenAI agents SDK |
| 20 | +# logging.basicConfig( |
| 21 | +# level=logging.INFO, format="%(asctime)s - %(name)s - %(levelname)s - %(message)s" |
| 22 | +# ) |
| 23 | +# logger.logger.setLevel(logging.ERROR) |
| 24 | + |
| 25 | + |
| 26 | +@function_tool |
| 27 | +def get_weather(city: str) -> str: |
| 28 | + """Get the weather in a city.""" |
| 29 | + return f"The weather in {city} is sunny." |
| 30 | + |
| 31 | + |
| 32 | +agent = RealtimeAgent( |
| 33 | + name="Assistant", |
| 34 | + instructions="You always greet the user with 'Top of the morning to you'.", |
| 35 | + tools=[get_weather], |
| 36 | +) |
| 37 | + |
| 38 | + |
| 39 | +def _truncate_str(s: str, max_length: int) -> str: |
| 40 | + if len(s) > max_length: |
| 41 | + return s[:max_length] + "..." |
| 42 | + return s |
| 43 | + |
| 44 | + |
| 45 | +class NoUIDemo: |
| 46 | + def __init__(self) -> None: |
| 47 | + self.session: RealtimeSession | None = None |
| 48 | + self.audio_stream: sd.InputStream | None = None |
| 49 | + self.audio_player: sd.OutputStream | None = None |
| 50 | + self.recording = False |
| 51 | + |
| 52 | + # Audio output state for callback system |
| 53 | + self.output_queue: queue.Queue[Any] = queue.Queue(maxsize=10) # Buffer more chunks |
| 54 | + self.interrupt_event = threading.Event() |
| 55 | + self.current_audio_chunk: np.ndarray | None = None # type: ignore |
| 56 | + self.chunk_position = 0 |
| 57 | + |
| 58 | + def _output_callback(self, outdata, frames: int, time, status) -> None: |
| 59 | + """Callback for audio output - handles continuous audio stream from server.""" |
| 60 | + if status: |
| 61 | + print(f"Output callback status: {status}") |
| 62 | + |
| 63 | + # Check if we should clear the queue due to interrupt |
| 64 | + if self.interrupt_event.is_set(): |
| 65 | + # Clear the queue and current chunk state |
| 66 | + while not self.output_queue.empty(): |
| 67 | + try: |
| 68 | + self.output_queue.get_nowait() |
| 69 | + except queue.Empty: |
| 70 | + break |
| 71 | + self.current_audio_chunk = None |
| 72 | + self.chunk_position = 0 |
| 73 | + self.interrupt_event.clear() |
| 74 | + outdata.fill(0) |
| 75 | + return |
| 76 | + |
| 77 | + # Fill output buffer from queue and current chunk |
| 78 | + outdata.fill(0) # Start with silence |
| 79 | + samples_filled = 0 |
| 80 | + |
| 81 | + while samples_filled < len(outdata): |
| 82 | + # If we don't have a current chunk, try to get one from queue |
| 83 | + if self.current_audio_chunk is None: |
| 84 | + try: |
| 85 | + self.current_audio_chunk = self.output_queue.get_nowait() |
| 86 | + self.chunk_position = 0 |
| 87 | + except queue.Empty: |
| 88 | + # No more audio data available - this causes choppiness |
| 89 | + # Uncomment next line to debug underruns: |
| 90 | + # print(f"Audio underrun: {samples_filled}/{len(outdata)} samples filled") |
| 91 | + break |
| 92 | + |
| 93 | + # Copy data from current chunk to output buffer |
| 94 | + remaining_output = len(outdata) - samples_filled |
| 95 | + remaining_chunk = len(self.current_audio_chunk) - self.chunk_position |
| 96 | + samples_to_copy = min(remaining_output, remaining_chunk) |
| 97 | + |
| 98 | + if samples_to_copy > 0: |
| 99 | + chunk_data = self.current_audio_chunk[ |
| 100 | + self.chunk_position : self.chunk_position + samples_to_copy |
| 101 | + ] |
| 102 | + # More efficient: direct assignment for mono audio instead of reshape |
| 103 | + outdata[samples_filled : samples_filled + samples_to_copy, 0] = chunk_data |
| 104 | + samples_filled += samples_to_copy |
| 105 | + self.chunk_position += samples_to_copy |
| 106 | + |
| 107 | + # If we've used up the entire chunk, reset for next iteration |
| 108 | + if self.chunk_position >= len(self.current_audio_chunk): |
| 109 | + self.current_audio_chunk = None |
| 110 | + self.chunk_position = 0 |
| 111 | + |
| 112 | + async def run(self) -> None: |
| 113 | + print("Connecting, may take a few seconds...") |
| 114 | + |
| 115 | + # Initialize audio player with callback |
| 116 | + chunk_size = int(SAMPLE_RATE * CHUNK_LENGTH_S) |
| 117 | + self.audio_player = sd.OutputStream( |
| 118 | + channels=CHANNELS, |
| 119 | + samplerate=SAMPLE_RATE, |
| 120 | + dtype=FORMAT, |
| 121 | + callback=self._output_callback, |
| 122 | + blocksize=chunk_size, # Match our chunk timing for better alignment |
| 123 | + ) |
| 124 | + self.audio_player.start() |
| 125 | + |
| 126 | + try: |
| 127 | + runner = RealtimeRunner(agent) |
| 128 | + async with await runner.run() as session: |
| 129 | + self.session = session |
| 130 | + print("Connected. Starting audio recording...") |
| 131 | + |
| 132 | + # Start audio recording |
| 133 | + await self.start_audio_recording() |
| 134 | + print("Audio recording started. You can start speaking - expect lots of logs!") |
| 135 | + |
| 136 | + # Process session events |
| 137 | + async for event in session: |
| 138 | + await self._on_event(event) |
| 139 | + |
| 140 | + finally: |
| 141 | + # Clean up audio player |
| 142 | + if self.audio_player and self.audio_player.active: |
| 143 | + self.audio_player.stop() |
| 144 | + if self.audio_player: |
| 145 | + self.audio_player.close() |
| 146 | + |
| 147 | + print("Session ended") |
| 148 | + |
| 149 | + async def start_audio_recording(self) -> None: |
| 150 | + """Start recording audio from the microphone.""" |
| 151 | + # Set up audio input stream |
| 152 | + self.audio_stream = sd.InputStream( |
| 153 | + channels=CHANNELS, |
| 154 | + samplerate=SAMPLE_RATE, |
| 155 | + dtype=FORMAT, |
| 156 | + ) |
| 157 | + |
| 158 | + self.audio_stream.start() |
| 159 | + self.recording = True |
| 160 | + |
| 161 | + # Start audio capture task |
| 162 | + asyncio.create_task(self.capture_audio()) |
| 163 | + |
| 164 | + async def capture_audio(self) -> None: |
| 165 | + """Capture audio from the microphone and send to the session.""" |
| 166 | + if not self.audio_stream or not self.session: |
| 167 | + return |
| 168 | + |
| 169 | + # Buffer size in samples |
| 170 | + read_size = int(SAMPLE_RATE * CHUNK_LENGTH_S) |
| 171 | + |
| 172 | + try: |
| 173 | + while self.recording: |
| 174 | + # Check if there's enough data to read |
| 175 | + if self.audio_stream.read_available < read_size: |
| 176 | + await asyncio.sleep(0.01) |
| 177 | + continue |
| 178 | + |
| 179 | + # Read audio data |
| 180 | + data, _ = self.audio_stream.read(read_size) |
| 181 | + |
| 182 | + # Convert numpy array to bytes |
| 183 | + audio_bytes = data.tobytes() |
| 184 | + |
| 185 | + # Send audio to session |
| 186 | + await self.session.send_audio(audio_bytes) |
| 187 | + |
| 188 | + # Yield control back to event loop |
| 189 | + await asyncio.sleep(0) |
| 190 | + |
| 191 | + except Exception as e: |
| 192 | + print(f"Audio capture error: {e}") |
| 193 | + finally: |
| 194 | + if self.audio_stream and self.audio_stream.active: |
| 195 | + self.audio_stream.stop() |
| 196 | + if self.audio_stream: |
| 197 | + self.audio_stream.close() |
| 198 | + |
| 199 | + async def _on_event(self, event: RealtimeSessionEvent) -> None: |
| 200 | + """Handle session events.""" |
| 201 | + try: |
| 202 | + if event.type == "agent_start": |
| 203 | + print(f"Agent started: {event.agent.name}") |
| 204 | + elif event.type == "agent_end": |
| 205 | + print(f"Agent ended: {event.agent.name}") |
| 206 | + elif event.type == "handoff": |
| 207 | + print(f"Handoff from {event.from_agent.name} to {event.to_agent.name}") |
| 208 | + elif event.type == "tool_start": |
| 209 | + print(f"Tool started: {event.tool.name}") |
| 210 | + elif event.type == "tool_end": |
| 211 | + print(f"Tool ended: {event.tool.name}; output: {event.output}") |
| 212 | + elif event.type == "audio_end": |
| 213 | + print("Audio ended") |
| 214 | + elif event.type == "audio": |
| 215 | + # Enqueue audio for callback-based playback |
| 216 | + np_audio = np.frombuffer(event.audio.data, dtype=np.int16) |
| 217 | + try: |
| 218 | + self.output_queue.put_nowait(np_audio) |
| 219 | + except queue.Full: |
| 220 | + # Queue is full - only drop if we have significant backlog |
| 221 | + # This prevents aggressive dropping that could cause choppiness |
| 222 | + if self.output_queue.qsize() > 8: # Keep some buffer |
| 223 | + try: |
| 224 | + self.output_queue.get_nowait() |
| 225 | + self.output_queue.put_nowait(np_audio) |
| 226 | + except queue.Empty: |
| 227 | + pass |
| 228 | + # If queue isn't too full, just skip this chunk to avoid blocking |
| 229 | + elif event.type == "audio_interrupted": |
| 230 | + print("Audio interrupted") |
| 231 | + # Signal the output callback to clear its queue and state |
| 232 | + self.interrupt_event.set() |
| 233 | + elif event.type == "error": |
| 234 | + print(f"Error: {event.error}") |
| 235 | + elif event.type == "history_updated": |
| 236 | + pass # Skip these frequent events |
| 237 | + elif event.type == "history_added": |
| 238 | + pass # Skip these frequent events |
| 239 | + elif event.type == "raw_model_event": |
| 240 | + print(f"Raw model event: {_truncate_str(str(event.data), 50)}") |
| 241 | + else: |
| 242 | + print(f"Unknown event type: {event.type}") |
| 243 | + except Exception as e: |
| 244 | + print(f"Error processing event: {_truncate_str(str(e), 50)}") |
| 245 | + |
| 246 | + |
| 247 | +if __name__ == "__main__": |
| 248 | + demo = NoUIDemo() |
| 249 | + try: |
| 250 | + asyncio.run(demo.run()) |
| 251 | + except KeyboardInterrupt: |
| 252 | + print("\nExiting...") |
| 253 | + sys.exit(0) |
0 commit comments