Does VAD is really working ?

[dilerbatu]

Hi I have a audio data like this:

audio = np.frombuffer(audio_msg.data, dtype=np.int16)

self._oww = Model(
                wakeword_models=[self._OWW_MODEL_PATH],
                inference_framework=self._OWW_INFERENCE_FRAMEWORK,
                vad_threshold=1
            )

and basically I run oww like this:

for mdl in self._oww.prediction_buffer.keys():
                    scores = list(self._oww.prediction_buffer[mdl])
                    curr_score = format(scores[-1], '.20f').replace("-", "")
                    print("Current score:", curr_score, end='\r', flush=True)

                    score = scores[-1]
                    if score >= self._OWW_THRESHOLD:
                        print('Detected wakeword by openwakeword with the score of', score)
                        break

My chunk size is 512. How does it work ? My VAD threshold is 1. It should not create score at all but when I say my wakeword it recognize it.

[dscripka]

It's difficult to determine what might be happening without seeing the full code, but you can start by inspecting the VAD object prediction buffer. This will show what the VAD model is producing, and then you can review the logic here and see if everything is functioning as expected.