Skip to main content
Keyterm prompting improves transcription accuracy for important terms and phrases by biasing the model toward transcribing them. It’s most useful for important or uncommon words and phrases, such as domain-specific terminology, brand or product names, and rare or invented words.
Pass one or more terms with the keyterm query parameter when you open the WebSocket connection:
To pass multiple terms, repeat the keyterm parameter, one per term:
For a multi-word phrase, keep it in a single keyterm value and encode spaces as %20:
You can pass up to 100 keyterms per connection, up to a total of 1200 characters across all keyterms.
Keyterms apply for the entire connection and are set only when you open it. You cannot add or change keyterms mid-stream.
Keyterm prompting works on both the Realtime STT and Realtime STT (Auto) endpoints.

​
SDK usage

The Cartesia Python and TypeScript/JavaScript SDKs both accept a keyterm argument on the STT websocket clients. Pass a list of strings — one entry per term or phrase. The SDK forwards each value as a repeated keyterm query parameter, so you do not need to build the URL yourself.
The same keyterm argument is available on client.stt.manual_finalize.websocket(...) and on the async equivalents (AsyncCartesia).

​
Best practices

Consider the following to achieve the best results:
  • Use keyterms for important or uncommon words and phrases, such as domain-specific terminology, brand or product names, and rare or invented words
  • Avoid generic, common words like “the” or “and”
  • Avoid using too many keyterms and prioritize the ones that matter most
  • Use exact spelling and capitalization for each keyterm