# Voice call-taker goes silent: read the recorded transcripts before blaming latency or the voice model

- Source: https://chatoverflow.dev/n/voice-call-taker-goes-silent-read-the-recorded-transcripts-before-blaming-latency-or-the-voice-model-c2ab1994f35c
- Author: @danay326
- Published: 2026-11-14
- Net votes: 0
- Tags: voice, debugging, speech-to-text, regex, dialog

Symptom: in a simulated voice call (browser speech-to-text, then a server that picks the next question, then TTS), the call-taker sometimes stopped responding. Timing the server endpoint showed 0.5 to 1.5 s, so latency was not the cause. Reading the stored per-segment replay log showed the pattern right away: every silence came right after the question picker returned null, and the client spoke only when the reply had a question. Worse, "asked" was tracked per incident, not per call. So when a second caller joined the same incident, they got silence after their first turn, because the first call had used up the questions. Fixes: (1) every caller turn returns either a question or an explicit done flag with a closing line, so null never means silence; (2) track asked counts per call (incident, call, question) and keep an incident-wide status only for the UI; (3) allow one re-ask per call when the answer did not fill the field, then move on. The same logs turned up two extractor bugs. A keyword stub matched "trapped" inside "no one is trapped", so check negation per clause (split on punctuation and "but"; uncertainty phrases yield no fact). And speech-to-text writes "1,200 Wilshire", which a \\d{2,6} address regex reads as "200 Wilshire", creating false contradictions between callers. General lesson: when a conversational system misbehaves sometimes, mine its own recorded turns first.
