Recognition hints
The words and phrases the recognition engine is told to expect, their limits, and the medical vocabulary switch.
Add the proper nouns your callers actually say: business names, suburbs, product codes, people's names. If a suburb keeps coming back as gibberish in transcripts, it belongs here.
Karrinyup
Joondalup
Dr Nguyen
XR-450 filterThe list lives on the Hearing page of the agent editor.
Hints are not versioned
Hints apply across every version of the agent, not just the draft, and the card is marked accordingly. That is the exception to the rest of the Hearing settings, which reach a call only when you publish.
Limits
- Up to 100 hints, each up to 100 characters.
- Each engine also bounds the list as a whole. The editor counts the remaining budget for the engine you have selected.
- Ink Whisper ignores hints entirely. Selecting it disables the input.
Going over budget does not block publishing, but it is flagged as a risk. The engine drops whatever does not fit, so hints at the end of the list quietly stop working.
Hints against pronunciation
Recognition hints fix what the agent hears. The pronunciation list on the Voice page fixes what it says. A clinic name or a product range usually needs both.
Medical vocabulary
An agent can be switched from general to medical vocabulary, which biases recognition towards clinical language: medication names, anatomy and procedures. The control sits on the Voice page because, like recognition hints, it applies to the whole agent rather than to one version.
Medical vocabulary works by swapping the Deepgram Nova model for its clinical variant. Flux and the Cartesia engines have no clinical variant, so on those it does nothing. Selecting medical vocabulary alongside an incompatible engine is flagged as a publish risk: the call still runs on the general model, without the boost you asked for.