Realtime: move demo audio to separate thread

Author: rm-openai

examples/realtime/demo.py

@@ -93,7 +93,8 @@ async def _on_event(self, event: RealtimeSessionEvent) -> None:
                 self.ui.add_transcript("Audio ended")
             elif event.type == "audio":
                 np_audio = np.frombuffer(event.audio.data, dtype=np.int16)
-                self.ui.play_audio(np_audio)
+                # Play audio in a separate thread to avoid blocking the event loop
+                await asyncio.to_thread(self.ui.play_audio, np_audio)
             elif event.type == "audio_interrupted":
                 self.ui.add_transcript("Audio interrupted")
             elif event.type == "error":

examples/realtime/no_ui_demo.py

@@ -1,5 +1,7 @@
 import asyncio
+import queue
 import sys
+import threading
 
 import numpy as np
 import sounddevice as sd
@@ -46,14 +48,77 @@ def __init__(self) -> None:
         self.audio_player: sd.OutputStream | None = None
         self.recording = False
 
+        # Audio output state for callback system
+        self.output_queue: queue.Queue[np.ndarray] = queue.Queue(maxsize=10)  # Buffer more chunks
+        self.interrupt_event = threading.Event()
+        self.current_audio_chunk: np.ndarray | None = None
+        self.chunk_position = 0
+
+    def _output_callback(self, outdata: np.ndarray, frames: int, time, status) -> None:
+        """Callback for audio output - handles continuous audio stream from server."""
+        if status:
+            print(f"Output callback status: {status}")
+
+        # Check if we should clear the queue due to interrupt
+        if self.interrupt_event.is_set():
+            # Clear the queue and current chunk state
+            while not self.output_queue.empty():
+                try:
+                    self.output_queue.get_nowait()
+                except queue.Empty:
+                    break
+            self.current_audio_chunk = None
+            self.chunk_position = 0
+            self.interrupt_event.clear()
+            outdata.fill(0)
+            return
+
+        # Fill output buffer from queue and current chunk
+        outdata.fill(0)  # Start with silence
+        samples_filled = 0
+
+        while samples_filled < len(outdata):
+            # If we don't have a current chunk, try to get one from queue
+            if self.current_audio_chunk is None:
+                try:
+                    self.current_audio_chunk = self.output_queue.get_nowait()
+                    self.chunk_position = 0
+                except queue.Empty:
+                    # No more audio data available - this causes choppiness
+                    # Uncomment next line to debug underruns:
+                    # print(f"Audio underrun: {samples_filled}/{len(outdata)} samples filled")
+                    break
+
+            # Copy data from current chunk to output buffer
+            remaining_output = len(outdata) - samples_filled
+            remaining_chunk = len(self.current_audio_chunk) - self.chunk_position
+            samples_to_copy = min(remaining_output, remaining_chunk)
+
+            if samples_to_copy > 0:
+                chunk_data = self.current_audio_chunk[
+                    self.chunk_position : self.chunk_position + samples_to_copy
+                ]
+                # More efficient: direct assignment for mono audio instead of reshape
+                outdata[samples_filled : samples_filled + samples_to_copy, 0] = chunk_data
+                samples_filled += samples_to_copy
+                self.chunk_position += samples_to_copy
+
+                # If we've used up the entire chunk, reset for next iteration
+                if self.chunk_position >= len(self.current_audio_chunk):
+                    self.current_audio_chunk = None
+                    self.chunk_position = 0
+
     async def run(self) -> None:
         print("Connecting, may take a few seconds...")
 
-        # Initialize audio player
+        # Initialize audio player with callback
+        chunk_size = int(SAMPLE_RATE * CHUNK_LENGTH_S)
         self.audio_player = sd.OutputStream(
             channels=CHANNELS,
             samplerate=SAMPLE_RATE,
             dtype=FORMAT,
+            callback=self._output_callback,
+            blocksize=chunk_size,  # Match our chunk timing for better alignment
         )
         self.audio_player.start()
 
@@ -146,15 +211,24 @@ async def _on_event(self, event: RealtimeSessionEvent) -> None:
             elif event.type == "audio_end":
                 print("Audio ended")
             elif event.type == "audio":
-                # Play audio through speakers
+                # Enqueue audio for callback-based playback
                 np_audio = np.frombuffer(event.audio.data, dtype=np.int16)
-                if self.audio_player:
-                    try:
-                        self.audio_player.write(np_audio)
-                    except Exception as e:
-                        print(f"Audio playback error: {e}")
+                try:
+                    self.output_queue.put_nowait(np_audio)
+                except queue.Full:
+                    # Queue is full - only drop if we have significant backlog
+                    # This prevents aggressive dropping that could cause choppiness
+                    if self.output_queue.qsize() > 8:  # Keep some buffer
+                        try:
+                            self.output_queue.get_nowait()
+                            self.output_queue.put_nowait(np_audio)
+                        except queue.Empty:
+                            pass
+                    # If queue isn't too full, just skip this chunk to avoid blocking
             elif event.type == "audio_interrupted":
                 print("Audio interrupted")
+                # Signal the output callback to clear its queue and state
+                self.interrupt_event.set()
             elif event.type == "error":
                 print(f"Error: {event.error}")
             elif event.type == "history_updated":

pyproject.toml

@@ -7,7 +7,7 @@ requires-python = ">=3.9"
 license = "MIT"
 authors = [{ name = "OpenAI", email = "[email protected]" }]
 dependencies = [
-    "openai>=1.96.0, <2",
+    "openai>=1.96.1, <2",
     "pydantic>=2.10, <3",
     "griffe>=1.5.6, <2",
     "typing-extensions>=4.12.2, <5",

src/agents/realtime/items.py

@@ -30,6 +30,15 @@ class AssistantText(BaseModel):
     model_config = ConfigDict(extra="allow")
 
 
+class AssistantAudio(BaseModel):
+    type: Literal["audio"] = "audio"
+    audio: str | None = None
+    transcript: str | None = None
+
+    # Allow extra data
+    model_config = ConfigDict(extra="allow")
+
+
 class SystemMessageItem(BaseModel):
     item_id: str
     previous_item_id: str | None = None
@@ -58,7 +67,7 @@ class AssistantMessageItem(BaseModel):
     type: Literal["message"] = "message"
     role: Literal["assistant"] = "assistant"
     status: Literal["in_progress", "completed", "incomplete"] | None = None
-    content: list[AssistantText]
+    content: list[Annotated[AssistantText | AssistantAudio, Field(discriminator="type")]]
 
     # Allow extra data
     model_config = ConfigDict(extra="allow")

src/agents/realtime/openai_realtime.py

@@ -364,7 +364,7 @@ async def _handle_output_item(self, item: ConversationItem) -> None:
                     "item_id": item.id or "",
                     "type": item.type,
                     "role": item.role,
-                    "content": item.content,
+                    "content": [content.model_dump() for content in item.content],
                     "status": "in_progress",
                 }
             )