[johnmatthewtennant]

Summary

Berd Call needs recordings and repeatable measurements to tune speech detection from daily use. This adds device-specific listening and interruption profiles, opt-in local call capture, offline replay and evaluation of detection settings, and fixes to the speech hold that keeps agents from talking over the user.

Capture. berd-call start --capture DIR records microphone audio, sent agent audio, timestamped protocol events, per-frame VAD decisions, recognizer generations, the default input and output devices (including changes mid-call), and each Apple final's audio end with how many speech turns it resolved.

Replay and evaluation. berd-call vad-replay events CAPTURE_DIR measures request waiting against VAD and pending recognition, and the delay from the end of detected speech to each final transcript and each hold release. convert, extract and evaluate turn captures into grouped fixtures and compare static candidates against a baseline on held-out sessions. learning_evaluation::evaluate_corpus exposes an EligibleCandidate only for corpora with exact device evidence; today's captures remain diagnostic (trial_ineligible), so nothing is promoted automatically. detection_profiles::prepare_trial and finish_trial keep the saved baseline until a trial passes on the same devices and settings.

Speech hold. Agent speech waits while the user speaks and while Apple still owes a final. Previously each VAD turn added a pending credit and each final resolved one, so a sentence split by pauses left credits outstanding and nearly every hold ran the full pending-utterance timeout (55 of 57 holds across two recorded calls). The native bridge now passes each final's audio end, and a final resolves exactly the ended turns it covers; a final for an earlier phrase of a still-pending turn resolves nothing. Apple can take seven seconds or more to finalize a second sentence. So the bridge also reports Apple's interim results, and while its latest result is interim words rather than a final, the hold lasts up to ten seconds past the timeout. Noise Apple never transcribes keeps the timeout. When the hold does expire, recognition keeps running and late finals are still published. Empty recognition results no longer raise detection thresholds.

Related issue

Consolidates #421, #422 and #424. No related issue found.

Testing

• cargo test -p berd-call passes locally on macOS at e173655 (400 library and 284 binary tests plus every integration target), including regressions for multi-turn and split-phrase finals, holds that wait for interim words and holds that keep their timeout once a final follows them, cancellation while live playback waits for provider audio (with buffered tails of 1 and 4095 samples), and capture route recording.
• cargo test -p berd-call --lib final_end_is_measured_on_the_pushed_audio_timeline -- --ignored runs real on-device SpeechTranscriber over say output with 1 s and 3 s of leading silence. Final ends shift by the lead within 40 ms and land about 0.3 s after the last sound, which confirms finals report the same timeline the session pushes.
• On recorded calls, vad-replay events shows finals arriving a median 1.8 to 2.6 s and at p90 3.1 to 4.1 s after speech ends, while holds ran 5.0 s.
• A live call on 43efb0e had a median hold of 2.40 s, with 1 of 14 holds at the 5 s ceiling. In that one, the second sentence finalized 3.8 s after the hold released, and an agent's reply started and was cut off.
• Replaying that call's microphone audio in real time through e173655: no final arrived after its hold released, and that hold lasted 8.7 s. Replaying a 20-minute call: median hold 2.73 s, max 5.06 s, the same as before the interim-words change.

Automatic promotion from captured calls is not implemented: captured corpora lack verified playback provenance and labels, so evaluate_corpus returns no eligible candidate.