What is voice activity detection doing in the pipeline? Audio arrives in chunks. Each chunk gets a speech-or-not label. Consecutive speech frames become an utterance; trailing silence becomes an endpoint that can trigger finalization. If VAD is too timid, you clip the first word. If it is too eager, fans and keyboards keep the recognizer busy. That is why a neural VAD, trained to ignore non-speech, usually beats a raw energy gate in an open office, while energy VAD remains useful when the neural file is missing.
Silero VAD is a widely used open model for this job: small enough to run on-device, accurate enough to replace many energy heuristics. It is not a speech-to-text model; it only answers 'is someone talking?' Endpointing for dictation still combines VAD with your push-to-talk or toggle key. A hold-to-talk session can ignore VAD for start, then use it to trim. Always-on capture without a key is a different product, and a much sharper privacy question, even with a good VAD in the loop.