Reduce live caller latency and improve reliability

- Replace per-callback async task spawning with persistent queue-based sender
- Buffer host mic to 60ms chunks (was 21ms) to reduce WebSocket frame rate
- Reduce server ring buffer prebuffer from 150ms to 80ms
- Reduce browser playback jitter buffer from 150ms to 100ms

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

backend/main.py

@@ -872,8 +872,14 @@ async def caller_audio_stream(websocket: WebSocket):
 
 # --- Host Audio Broadcast ---
 
-async def _broadcast_host_audio(pcm_bytes: bytes):
+_host_audio_queue: asyncio.Queue = None
-    """Send host mic audio to all active real callers"""
+_host_audio_task: asyncio.Task = None
+
+
+async def _host_audio_sender():
+    """Persistent task that drains audio queue and sends to callers"""
+    while True:
+        pcm_bytes = await _host_audio_queue.get()
         for caller_id in list(caller_service.active_calls.keys()):
             try:
                 await caller_service.send_audio_to_caller(caller_id, pcm_bytes, 16000)
@@ -881,16 +887,23 @@ async def _broadcast_host_audio(pcm_bytes: bytes):
                 pass
 
 
+def _start_host_audio_sender():
+    """Start the persistent host audio sender task"""
+    global _host_audio_queue, _host_audio_task
+    if _host_audio_queue is None:
+        _host_audio_queue = asyncio.Queue(maxsize=100)
+    if _host_audio_task is None or _host_audio_task.done():
+        _host_audio_task = asyncio.create_task(_host_audio_sender())
+
+
 def _host_audio_sync_callback(pcm_bytes: bytes):
-    """Sync wrapper to schedule async broadcast from audio thread"""
+    """Sync callback from audio thread — push to queue for async sending"""
+    if _host_audio_queue is None:
+        return
     try:
-        loop = asyncio.get_event_loop()
+        _host_audio_queue.put_nowait(pcm_bytes)
-        if loop.is_running():
+    except asyncio.QueueFull:
-            loop.call_soon_threadsafe(
+        pass  # Drop frame rather than block
-                asyncio.ensure_future, _broadcast_host_audio(pcm_bytes)
-            )
-    except Exception:
-        pass
 
 
 # --- Queue Endpoints ---
@@ -920,6 +933,7 @@ async def take_call_from_queue(caller_id: str):
 
     # Start host mic streaming if this is the first real caller
     if len(caller_service.active_calls) == 1:
+        _start_host_audio_sender()
         audio_service.start_host_stream(_host_audio_sync_callback)
 
     return {

backend/services/audio.py

@@ -350,10 +350,10 @@ class AudioService:
         self._live_caller_num_channels = num_channels
         self._live_caller_channel_idx = channel_idx
 
-        # Ring buffer: 3 seconds capacity, 150ms pre-buffer before playback starts
+        # Ring buffer: 3 seconds capacity, 80ms pre-buffer before playback starts
         ring_size = int(device_sr * 3)
         ring = np.zeros(ring_size, dtype=np.float32)
-        prebuffer_samples = int(device_sr * 0.15)
+        prebuffer_samples = int(device_sr * 0.08)
         # Mutable state shared between writer (main thread) and reader (audio callback)
         # CPython GIL makes individual int reads/writes atomic
         state = {"write_pos": 0, "read_pos": 0, "avail": 0, "started": False}
@@ -459,6 +459,11 @@ class AudioService:
             record_channel = min(self.input_channel, max_channels) - 1
             step = max(1, int(device_sr / 16000))
 
+            # Buffer host mic to send ~60ms chunks instead of tiny 21ms ones
+            host_accum = []
+            host_accum_samples = [0]
+            send_threshold = 960  # 60ms at 16kHz
+
             def callback(indata, frames, time_info, status):
                 # Capture for push-to-talk recording if active
                 if self._recording:
@@ -470,7 +475,15 @@ class AudioService:
                 # Simple decimation to ~16kHz
                 if step > 1:
                     mono = mono[::step]
-                pcm = (mono * 32767).astype(np.int16).tobytes()
+
+                host_accum.append(mono.copy())
+                host_accum_samples[0] += len(mono)
+
+                if host_accum_samples[0] >= send_threshold:
+                    combined = np.concatenate(host_accum)
+                    pcm = (combined * 32767).astype(np.int16).tobytes()
+                    host_accum.clear()
+                    host_accum_samples[0] = 0
                     self._host_send_callback(pcm)
 
             self._host_device_sr = device_sr

frontend/call-in.html

@@ -150,6 +150,6 @@
         </div>
     </div>
 
-    <script src="/js/call-in.js?v=3"></script>
+    <script src="/js/call-in.js?v=4"></script>
 </body>
 </html>

frontend/js/call-in.js

@@ -84,7 +84,7 @@ class PlaybackProcessor extends AudioWorkletProcessor {
         this.readPos = 0;
         this.available = 0;
         this.started = false;
-        this.jitterMs = 150; // buffer 150ms before starting playback
+        this.jitterMs = 100; // buffer 100ms before starting playback
         this.jitterSamples = Math.floor(16000 * this.jitterMs / 1000);
 
         this.port.onmessage = (e) => {