[Kayvan-Zahiri]

Summary

• VADAnalyzer checked start_secs and stop_secs once per call rather than once per analysis frame. _run_analyzer drains the whole buffer in a while loop, but both threshold checks sat after it, so a buffer holding several analysis frames could cross a threshold and fall back out of it before either check ran. A confidence dip mid-utterance then wiped the accumulated start count and speech that had already qualified stayed STARTING, with no on_speech_started for the turn. The checks now run inside the loop, which makes one iteration behave exactly like one frame delivered on its own. Chunk size no longer changes the state the analyzer settles on.

Bundled transports use 10 to 20 ms chunks and are unaffected. The WebSocket transports take whatever chunk size the client sends, and the while loop was added in #2184 because create_stream_resampler can emit larger buffers than expected.

Testing

uv run pytest tests/test_vad_analyzer.py tests/test_vad_controller.py tests/test_vad_processor.py

Reverting the source change fails the new analyzer tests at 4 and 12 frames per chunk with VADState.STARTING != VADState.SPEAKING.

I also swept the neighboring inputs. Across 9,660 comparisons of the same confidence script delivered at 2 to 24 analysis frames per chunk, over seven sample rate and threshold combinations including 8 kHz and a 10 ms analysis window, the settled state disagreed with frame by frame delivery 771 times before and 0 times after. Sub-frame and single-frame delivery, the path every bundled transport takes, is unchanged, as is the real Silero model at 16 kHz and 8 kHz, start_secs=0, thresholds that are never reached, the volume gate, empty buffers, and set_params mid-stream. Full suite failure set is identical to main on this machine.

One thing this does not fix: a buffer carrying an entire turn still returns only the state it ends on, so a caller reading one state per call can still miss transitions that begin and end inside one chunk. That is the shape of analyze_audio rather than the threshold placement, and it seemed like a separate conversation.

(This originally also wired speech_activity_period into _handle_audio; #5805 merged that same change on 9/16, so it's been dropped here.)

[Kayvan-Zahiri]

Rebased onto main to clear the conflict. #5805 landed the same _maybe_speech_activity() change as my second commit, so I dropped that commit and its changelog entry; this PR is now just the per-frame threshold fix. On current main, 5 of the 7 new tests in tests/test_vad_analyzer.py fail. On this branch the analyzer, controller and processor VAD tests all pass (18).