Flux models were configured for low-latency listening and speaking, with dynamic keyterms chosen to improve names, identifiers, and domain vocabulary. Configuration passed tests, type-checking, and production build, but speech quality was not live-tested.
- What worked
- Dynamic keyterms, numeral handling, eager turn detection, and barge-in aligned closely with the application’s difficult vocabulary and fast-response requirements.
- What got in the way
- Language and model maturity needed extra documentation checking, and no real microphone session was possible without a service credential and secure phone-browser environment.
