A clerk's office phone tree uses speech-to-text and text-to-speech but must transfer to a human when recognition confidence is low. Which safeguard should they configure?

The clerk’s office phone tree uses speech-to-text and text-to-speech but must transfer to a human when recognition confidence is low. Configure confidence-based human handoff when speech recognition certainty is low so callers are not trapped by a guessing bot. In municipal clerk services—licenses, records, court dates—wrong automated answers create legal and citizen-trust risk; low-confidence STT should escalate.

Configuring confidence-based human handoff when speech recognition model accuracy is low watches the wrong level, because overall model accuracy does not show whether this caller’s request was recognized correctly, so a misunderstood caller can still be trapped by the bot. Configuring confidence-based human handoff when speech recognition audio sampling is low triggers on capture sampling rather than recognition confidence, so a misrecognized request at normal sampling still goes through the bot without a reliable path to a person. Configuring confidence-based human handoff when speech recognition phrase latency is low ties the transfer to recognition speed rather than recognition confidence, so a slow but misunderstood caller can still be trapped by the bot while fast, correctly recognized calls would hand off needlessly. Those options all use the wrong trigger — model accuracy, audio sampling, or phrase latency — instead of speech recognition certainty for the real-time human contact the caller needs right now.

Exam caveat: confidence thresholds are an operational safety pattern for speech agents; they are not the same as model-accuracy, audio-sampling, or phrase-latency routing. When stems mention low recognition confidence, choose human handoff rather than “try harder” generation.

Operational check: set an STT confidence threshold from labeled calls, route below-threshold calls to clerks with the partial transcript attached, measure containment versus handoff rates, and review invent-attempts as critical defects. Publish a clear “press zero for a person” path even when confidence is high, for accessibility and trust.