Plan 3: openwakeword wrapper with threshold + cooldown

client/wakeword.py

@@ -0,0 +1,47 @@
+"""openwakeword wrapper for the Pi client.
+
+Loads the stock `alexa` model. Each call to `predict()` takes a 1280-sample
+16 kHz mono int16 frame (produced by `client.audio.resample_24k_to_16k`).
+Returns True at most once per cooldown window.
+"""
+import time
+
+import numpy as np
+
+from client.log import get_logger
+
+_log = get_logger("wakeword")
+
+WAKEWORD = "alexa"
+DEFAULT_THRESHOLD = 0.5
+DEFAULT_COOLDOWN_S = 1.5
+
+
+class WakewordDetector:
+    def __init__(self, *, threshold: float = DEFAULT_THRESHOLD,
+                 cooldown_s: float = DEFAULT_COOLDOWN_S) -> None:
+        # Lazy-import openwakeword to keep the test suite importable without it.
+        import openwakeword.utils
+        from openwakeword.model import Model
+
+        _log.info("downloading openwakeword models (idempotent)")
+        # No-arg form dodges the None-handling crash documented in findings.md §4.
+        openwakeword.utils.download_models()
+        _log.info("loading openwakeword model %r", WAKEWORD)
+        t0 = time.monotonic()
+        self._model = Model(wakeword_models=[WAKEWORD], inference_framework="onnx")
+        _log.info("openwakeword model loaded in %.2fs", time.monotonic() - t0)
+
+        self._threshold = threshold
+        self._cooldown_s = cooldown_s
+        self._last_fired = 0.0
+
+    def predict(self, frame_16k_int16: np.ndarray) -> bool:
+        scores = self._model.predict(frame_16k_int16)
+        score = float(scores.get(WAKEWORD, 0.0))
+        now = time.monotonic()
+        if score >= self._threshold and (now - self._last_fired) >= self._cooldown_s:
+            self._last_fired = now
+            _log.info("wakeword fired score=%.3f", score)
+            return True
+        return False