Simulates the telephone path: mic captured as 8 kHz PCM Int16 mono (the rate a real phone line delivers), then upsampled to 16 kHz. Audio streams continuously to the ASR service. Smart Turn now uses stateless HTTP — the client keeps a local 16 kHz rolling ring, and each ASR end event POSTs the last ≤ 8 s to /v1/analyze under a 100 ms latency ceiling (see docs/smart_turn/TRANSPORT_HTTP.md). Each Smart Turn probability is fused with six language-agnostic ASR signals — excess trailing silence, tail-confidence drop, fragment size, speech density, cross-segment cadence, and turn duration — into a final commit probability. See docs/smart_turn/SIGNALS.md.
longer SNR gate. When per-call
calibration is enabled, this remains the threshold during warm-up and
is replaced only after calibration locks.
enforce mode; off keeps the fallback longer SNR threshold for the entire session.
longer-event SNR threshold is calibrated.
A real longer SNR becomes trusted when the same VAD index
later produces a non-empty transcript—the livetest equivalent of that
index contributing input toward ChatCore. When enabled, the first
min_samples trusted values calculate and enforce one unbounded
bidirectional threshold, which may be as low or high as those values
require. Calibration then stops and later SNR values cannot change it.
Transcript/normal SNR is never added to this calibrator.
end events, then checks the speaker only on longer events when ASR SNR fails. On Stop the client calls /v1/end_session. See docs/ASR/speaker_gate/SPEAKER_GATE.md §Redesign.
/v1/enroll on every ASR end until the
server confirms enrolled. Once enrolled, a failed ASR SNR check on a
longer event sends the available audio to
/v1/check. All audio-processing decisions (accumulator threshold,
short-chunk floor, tail-trim) live on the server — see the
external_servers/speaker_gate_service env vars
(ENROLL_THRESHOLD_MS, CHECK_MIN_MS,
CHECK_MAX_MS). The server
returns raw similarity; the client compares against a
longer_similarity_threshold. The longer
slice uses the ASR start anchor with backward padding to recover speech
onset hidden by ASR event latency.