Skip to content

Keyterm boosting

Product names, drug names, jargon — the words that matter most are the ones generic models miss. Pass them as keyterms and the gateway translates to each vendor’s biasing mechanism (phrase lists, keyword boosts, keyterm prompts):

wss://api.speechrouter.ai/v1/listen?model=deepgram/nova-3&keyterms=SpeechRouter,Soniox,BYOK
sr.listen({ model: 'deepgram/nova-3', keyterms: ['SpeechRouter', 'Soniox', 'BYOK'] })

Works on batch too (-F keyterms=...).

Each vendor caps its list differently (AssemblyAI 100 terms, Azure 500, …) — the gateway forwards up to the vendor’s cap. Models without any biasing mechanism simply ignore the parameter; the provider pages mark support.

  • Boost rare terms only. Boosting common words hurts accuracy.
  • Casing is preserved where the vendor honors it (SpeechRouter, not speechrouter).
  • Need vendor-specific weighting knobs? Reach through with provider_params — e.g. Speechmatics additional_vocab with sounds_like, or AssemblyAI’s keyterms_prompt.