* refactor: run Silero VAD through transcribe.cpp instead of ONNX Moves VAD onto the engine we're consolidating on, so vad-rs is no longer an ONNX Runtime consumer (only transcribe-rs remains). Silero v6.2 GGUF replaces the v4 ONNX model; v4's architecture can't load in transcribe.cpp. VAD scores every 32 ms frame, so it runs in the app process rather than the engine worker, whose lifetime and serial request loop are tied to the ASR model. Only CPU backends are registered there, so no GPU driver code loads into the app. The 0.3 threshold and ms-based timing constants carry over; frame counts follow the model's 512-sample frames. * slim comments * restore Silero frame-size comment * comment * logging + minor review
1.2 MiB
1.2 MiB