Ctrlk
For the complete documentation index, see llms.txt. This page is also available as Markdown.

Deepgram's models struggle to identify names and other proper nouns, how can I get them to appear?

Depending on how commonly recognizable the named entity that you are trying to transcribe is (and the type of named entity that it is) Deepgram’s smart-formating feature or Keyword Boosting may be the best choice for addressing this need for your use case.

Keyword boosting works by programmatically specifying words that you would like the model in use to pay special attention to (or suppress) for recognition in transcription. At Deepgram, our keyword boosting feature works best on out-of-vocabulary (OOV) keywords (that the model has not been trained on). Product names, proper names, and industry-specific terms are all potentially good candidates for this feature. See the keywords documentation for additional details: https://developers.deepgram.com/docs/keywords.

Smart formatting helps make the transcript more readable and works well on most abbreviations (literary, alphabetisms, acronyms, and blends), alphanumerics, emails, currency, and other formatting but can fall short when it comes to less-common names and places, which is where custom model training can be very successful.

Custom model training is a process where we take audio, truth transcripts (provided by you and verified by us or generated by our human data labeling teams), and keywords (important to your use case) to create a customized ASR model that is tailored to your use case and vocabulary. Custom model training can greatly improve transcription accuracy overall, but especially in the case where our customers are looking to improve the recognition of people’s names and named entities.

Interested in custom model training? Reach out to speak to our Product Consultants, here: https://deepgram.com/contact-us/

For more information on smart formatting, keyword boosting, and custom model training, please refer to the associated resources below:

Last updated