Follow-up from the cloud review of #70 (P2).
DEFAULT_REFRACTORY_S = _NO_SPEECH_ONSET_FRAMES * FRAME_HOP_S (~1.6s) models the post-fire gap for the no-speech case: after a fire the live loop re-arms once it has consumed the request, at least ~1.6s when no speech follows.
But recording-computah.md asks users to make background.wav by talking or reading for ~2 min. On that continuous-speech audio, production does not re-arm after the 1.6s floor: once post-fire speech is seen it enters capture_request and re-arms only after endpoint silence or the max-request cap, both longer than 1.6s. So the sweep can count several false accepts across one sustained talk region where the live loop would still be capturing the first false turn. That inflates FA/hour and pushes the recommended threshold too high for continuous-speech backgrounds.
Workaround today: raise --refractory-s when the negatives are continuous speech. Possible fixes: derive the refractory gap from the request-endpointing model (endpoint-silence window + max-request cap) rather than the no-speech floor for spoken backgrounds, or document that --refractory-s should be set to the endpointing window for talk/read negatives.
Ref: review thread on #70, eval_wake_threshold.DEFAULT_REFRACTORY_S.
Follow-up from the cloud review of #70 (P2).
DEFAULT_REFRACTORY_S=_NO_SPEECH_ONSET_FRAMES * FRAME_HOP_S(~1.6s) models the post-fire gap for the no-speech case: after a fire the live loop re-arms once it has consumed the request, at least ~1.6s when no speech follows.But recording-computah.md asks users to make
background.wavby talking or reading for ~2 min. On that continuous-speech audio, production does not re-arm after the 1.6s floor: once post-fire speech is seen it enterscapture_requestand re-arms only after endpoint silence or the max-request cap, both longer than 1.6s. So the sweep can count several false accepts across one sustained talk region where the live loop would still be capturing the first false turn. That inflates FA/hour and pushes the recommended threshold too high for continuous-speech backgrounds.Workaround today: raise
--refractory-swhen the negatives are continuous speech. Possible fixes: derive the refractory gap from the request-endpointing model (endpoint-silence window + max-request cap) rather than the no-speech floor for spoken backgrounds, or document that--refractory-sshould be set to the endpointing window for talk/read negatives.Ref: review thread on #70,
eval_wake_threshold.DEFAULT_REFRACTORY_S.