⚡ Bolt: Add Parakeet model warmup

.jules/bolt.md

@@ -1,3 +1,7 @@
 ## 2024-02-18 - Pre-scaled Audio Feedback
 **Learning:** AudioFeedback volume scaling was applied during every playback, causing unnecessary numpy overhead and latency.
 **Action:** Pre-calculate scaled audio during `_load_and_cache` to minimize `_play_cached` latency (from ~1ms to ~0.04ms).
+
+## 2024-05-22 - ONNX Model Warmup
+**Learning:** The first inference pass with ONNX Runtime incurs a significant "cold start" penalty (buffer allocation, graph optimization) which impacts user perceived latency for the first dictation.
+**Action:** Implemented `ParakeetManager.warmup()` to run a dummy inference during application startup, shifting this cost away from the user interaction loop.

src/chirp/main.py

@@ -72,6 +72,7 @@ def __init__(self, *, verbose: bool = False) -> None:
                     model_dir=model_dir,
                     timeout=self.config.model_timeout,
                 )
+                self.parakeet.warmup()
         except ModelNotPreparedError as exc:
             self.logger.error(str(exc))
             raise SystemExit(1) from exc

src/chirp/parakeet_manager.py

@@ -78,6 +78,15 @@ def ensure_loaded(self):
                 self._model = self._load_model()
             return self._model
 
+    def warmup(self) -> None:
+        """Performs a dummy inference to initialize ONNX buffers."""
+        self._logger.debug("Warming up Parakeet model...")
+        dummy_audio = np.zeros(16_000, dtype=np.float32)
+        try:
+            self.transcribe(dummy_audio)
+        except Exception as exc:
+            self._logger.warning("Warmup failed: %s", exc)
+
     def _resolve_providers(self, key: str) -> Sequence[str]:
         normalized = key.lower()
         if normalized != "cpu":

tests/test_parakeet_manager.py

@@ -88,6 +88,31 @@ def test_transcribe_reloads_and_updates_time(self, mock_time, mock_onnx):
         manager._stop_monitor.set()
         time.sleep(0.05)
 
+    @patch("chirp.parakeet_manager.onnx_asr")
+    def test_warmup(self, mock_onnx):
+        """Test that warmup performs dummy inference."""
+        mock_model_instance = MagicMock()
+        mock_onnx.load_model.return_value = mock_model_instance
+
+        manager = ParakeetManager(
+            model_name="test",
+            quantization=None,
+            provider_key="cpu",
+            threads=1,
+            logger=self.logger,
+            model_dir=self.model_dir,
+            timeout=100.0,
+        )
+
+        manager.warmup()
+
+        # Verify recognize was called with silent audio
+        mock_model_instance.recognize.assert_called_once()
+        args, _ = mock_model_instance.recognize.call_args
+        audio_arg = args[0]
+        self.assertEqual(audio_arg.shape, (16000,))
+        self.assertTrue(np.all(audio_arg == 0))
+
     @patch("chirp.parakeet_manager.onnx_asr")
     def test_timeout_zero_disables_monitor(self, mock_onnx):
         """Test that timeout=0 disables the monitor thread."""