Make Silero VAD stateful across calls (carry LSTM state)

[danielbodart]

Problem

whisper.cpp's whisper_vad_detect_speech resets LSTM hidden/cell states on every call (ggml_backend_buffer_clear(vctx->buffer, 0) at whisper.cpp:5131). This is by design for whisper.cpp's one-shot file processing use case, but means our streaming usage (calling it repeatedly with 512-sample chunks) loses temporal context between calls.

The upstream Silero VAD model is designed to be stateful — LSTM state should carry across 512-sample chunks, just like TEN-VAD carries state across 256-sample hops.

Proposed change

Since we fork whisper.cpp, add a non-breaking way to skip the buffer clear:

• Option A: Add a bool reset_state parameter or flag to whisper_vad_detect_speech
• Option B: Add a separate whisper_vad_reset_state() function and remove the auto-reset from detect_speech
• Option C: Just remove the ggml_backend_buffer_clear line and let callers explicitly reset via a new function when needed

Also need to make VadBackend.reset() for Silero call the new reset function (currently a no-op).

Impact

• No change to our calling code — SileroVad.chunkProbS16 already calls once per chunk
• Probabilities should improve with temporal context
• Likely improves Silero's accuracy on our regression tests
• No upstream issue/PR exists for this (checked Feb 2026)

Priority

Low — TEN-VAD is our default and already stateful. This only affects --vad silero.

[danielbodart]

Fixed in 464ea41 — Make Silero VAD stateful across streaming chunks.

Uses whisper_vad_detect_speech_no_reset to carry LSTM state across chunks, with explicit reset between utterances. Regression results showed WER improvements across the board.

Upstream PR: ggml-org/whisper.cpp#3677 (open — we run on our fork in the meantime).

Default VAD was subsequently switched to Silero in d0a7a9c after it consistently outperformed TEN-VAD GGML on medium/long content.