We are releasing an updated keyword boosting feature (keywords={keyword}) with improved support for out of vocabulary words. Keywords can be boosted or suppressed with this feature.  Note, that this is not a persistent feature and every API call that requires keyword boosting or suppression needs to be send every time.

This feature is useful in two situations: (1) recognizing words in domain with a large vocabulary of domain-specific jargon and (2) a client application that depends on recognizing names, brands, or other relatively rare keywords with high accuracy.

We recommend using this feature in combination with a custom trained model from Deepgram. Deepgram will customize a model for your domain, ensuring that the vocabulary of the model fits your domain. Keyword boosting allows customers to add additional words rapidly (i.e. new brands, new menu items, etc) to augment a custom model’s domain-focused vocabulary.

To adjust the default strength of the keyword, assign a boost score to each supplied keyword like so — keywords=Deepgram:3. To suppress keywords, apply a negative boost score — keywords=Deepgram:-10. The simple form keywords=Deepgram is equivalent to keywords=Deepgram:1, and keywords=Deepgram:0 is equivalent to no boost at all (as if you never passed the keywords parameter in the first place). The boosts apply exponentially, so keywords=Deepgram:4 is a LOT more powerful than keywords=Deepgram:3.

To opt out of the new functionality, please add the query parameter keyword_boost=legacy to your request.

Read more about specifying keywords in the Deepgram Documentation.

Stop building work-arounds for STT systems that don't work.

Start FreeTalk to an expert