fix: reject near-silent audio to prevent Whisper hallucinations

[egsok]

Summary

• Add RMS energy gate (threshold 0.005) in the transcription pipeline — skips audio that passed VAD but is too quiet for real speech
• Add minimum speech duration check (100ms / 1600 samples) in audio manager — discards tiny VAD leakage fragments before they get zero-padded to 1.25s

Problem

When recording a few seconds of "silence", Whisper hallucinates text like "Subtitles by the Amara.org community". Real microphones pick up ambient noise that occasionally exceeds VAD's threshold, causing SmoothedVad to flush its prefill buffer (~8000+ samples of near-silence). The padding logic then inflates this to 1.25 seconds, and Whisper hallucinates on it.

How it works

Primary defense — RMS energy gate (transcription.rs): Computes RMS of the audio buffer and rejects anything below 0.005. Mic self-noise is ~0.0001–0.001, ambient noise that fools VAD is ~0.001–0.005, while whispered speech starts at ~0.01. This catches all near-silent audio regardless of how it got through VAD.

Secondary defense — min duration (audio.rs): Fragments shorter than 100ms (1600 samples at 16kHz) are discarded as VAD leakage — SmoothedVad's minimum real output is ~8000 samples. A cheap safety net for edge cases.

Test plan

• Record 2–3 seconds of silence → no transcription output, no hallucination
• Record a short word (~0.3s) → transcribes normally
• Record normal speech → identical behavior to before
• Debug mode (Ctrl+Shift+D): "Audio RMS ... below silence threshold" logged when recording silence

[cjpais]

I think this is probably a good change, I will test it and pull it in. May need to pull in beta testers too to get more feedback on it to ensure no unintended consequences. But I think if there is very low RMS there should be no big problem

[cjpais]

Hmmm.. I need to think more on this. I am not convinced that we can ship this broadly to everyone without consequences