⚡ Bolt: Vectorize frame energy calculation

.jules/bolt.md

@@ -0,0 +1,3 @@
+## 2025-04-11 - Vectorize audio processing loops
+**Learning:** Python `for` loops iterating over array slices are slow for audio framing tasks.
+**Action:** Use NumPy vectorization by reshaping the 1D audio array to 2D and computing aggregations like `np.mean(..., axis=1)`.

MANIFEST.in

@@ -48,6 +48,9 @@ recursive-include config *
 recursive-include k8s *
 recursive-include .github *.yml *.yaml *.md
 
+# Slower Whisper package explicitly included
+recursive-include slower_whisper *.py
+
 # Exclude compiled/temporary artifacts
 global-exclude __pycache__
 global-exclude *.py[cod]

config/Dockerfile

@@ -70,6 +70,7 @@ COPY pyproject.toml uv.lock README.md ./
 COPY transcription/ ./transcription/
 COPY scripts/ ./scripts/
 COPY integrations/ ./integrations/
+COPY slower_whisper/ ./slower_whisper/
 
 # Install Python dependencies using uv
 # ARG INSTALL_MODE controls which dependencies to install
@@ -128,6 +129,7 @@ COPY --from=builder /usr/local/bin /usr/local/bin
 COPY --chown=appuser:appuser transcription/ /app/transcription/
 COPY --chown=appuser:appuser scripts/ /app/scripts/
 COPY --chown=appuser:appuser integrations/ /app/integrations/
+COPY --chown=appuser:appuser slower_whisper/ /app/slower_whisper/
 COPY --chown=appuser:appuser pyproject.toml /app/
 
 # Create data directories

config/Dockerfile.gpu

@@ -91,6 +91,7 @@ COPY pyproject.toml uv.lock README.md ./
 COPY transcription/ ./transcription/
 COPY scripts/ ./scripts/
 COPY integrations/ ./integrations/
+COPY slower_whisper/ ./slower_whisper/
 
 # Install Python dependencies using uv
 # ARG INSTALL_MODE controls which dependencies to install

tests/test_streaming_asr.py

@@ -139,39 +139,6 @@ def test_pcm_bytes_to_float32_normalization(self):
         # Should be close to 1.0 (32767/32768)
         assert float32[0] == pytest.approx(32767 / 32768, rel=1e-4)
 
-    def test_calculate_energy_silence(self):
-        """Test energy calculation for silence."""
-        model = MockWhisperModel()
-        adapter = StreamingASRAdapter(model)
-
-        silence = np.zeros(1600, dtype=np.float32)
-        energy = adapter._calculate_energy(silence)
-
-        assert energy == 0.0
-
-    def test_calculate_energy_signal(self):
-        """Test energy calculation for non-zero signal."""
-        model = MockWhisperModel()
-        adapter = StreamingASRAdapter(model)
-
-        # Sine wave should have non-zero energy
-        t = np.linspace(0, 1, 16000, dtype=np.float32)
-        signal = 0.5 * np.sin(2 * np.pi * 440 * t)  # 440 Hz tone
-        energy = adapter._calculate_energy(signal)
-
-        assert energy > 0.0
-        assert energy < 1.0
-
-    def test_calculate_energy_empty(self):
-        """Test energy calculation for empty array."""
-        model = MockWhisperModel()
-        adapter = StreamingASRAdapter(model)
-
-        empty = np.array([], dtype=np.float32)
-        energy = adapter._calculate_energy(empty)
-
-        assert energy == 0.0
-
     @pytest.mark.asyncio
     async def test_ingest_audio_empty(self):
         """Test ingesting empty audio."""

transcription/streaming_asr.py

@@ -173,19 +173,6 @@ def _pcm_bytes_to_float32(self, pcm_bytes: bytes) -> np.ndarray:
 
         return float32_array
 
-    def _calculate_energy(self, audio: np.ndarray) -> float:
-        """Calculate RMS energy of audio segment.
-
-        Args:
-            audio: Float32 audio array.
-
-        Returns:
-            RMS energy value (0.0-1.0 for normalized audio).
-        """
-        if len(audio) == 0:
-            return 0.0
-        return float(np.sqrt(np.mean(audio**2)))
-
     def _detect_speech_frames(self, audio: np.ndarray, frame_size_ms: int = 30) -> list[bool]:
         """Detect speech in audio using frame-wise energy analysis.
 
@@ -199,13 +186,20 @@ def _detect_speech_frames(self, audio: np.ndarray, frame_size_ms: int = 30) -> l
         frame_size = int(self.config.sample_rate * frame_size_ms / 1000)
         num_frames = len(audio) // frame_size
 
-        speech_frames = []
-        for i in range(num_frames):
-            frame = audio[i * frame_size : (i + 1) * frame_size]
-            energy = self._calculate_energy(frame)
-            speech_frames.append(energy > self.config.vad_energy_threshold)
+        if num_frames == 0:
+            return []
+
+        # Truncate to exact multiple of frame_size
+        truncated_audio = audio[: num_frames * frame_size]
+
+        # Reshape into (num_frames, frame_size)
+        frames = truncated_audio.reshape((num_frames, frame_size))
+
+        # Calculate energy across frames: sqrt(mean(audio^2, axis=1))
+        energies = np.sqrt(np.mean(frames**2, axis=1))
 
-        return speech_frames
+        # Boolean array to list of bools
+        return [bool(e > self.config.vad_energy_threshold) for e in energies]
 
     def _process_vad(
         self,