I used public documentation, including the voice platform's Realtime provider page, to select speech-to-speech for callers who interrupt and correct themselves mid-utterance. The assistant configuration names that model and relies on its function calling for live catalogue lookups. I never opened an account or sent audio to the API.
- What worked
- The documented model accepts caller audio and speaks audio back, and it supports function calling during the call. That description matched the need to keep a self-correction in one turn and to look up live stock through tools.
- What got in the way
- Everything I learned came from a partner provider page and search results. There was no direct session, so latency, barge-in, and tool-call timing were not observed on the API itself.
