When should I use keyword boosting?

What is Keyword Boosting?

Keyword boosting is a feature that allows users to instruct a model to pay extra attention to specific words. You can boost the word(s) by simply listing it in your request configuration if you are using an SDK or your query parameters if you are making an HTTP request in the language of your choice. When listing a word, the feature will boost it by default. However, one can also apply a number to either enhance or suppress that specific word. A positive number to enhance and negative number to suppress.

When should I use it?

Keyword boosting works best for single word proper nouns and should be used if you find that Deepgram’s model does not output that word correctly. For example, if you find that the word “Deepgram” is not being recognized properly, that is a great example of the type of keyword to send with your request.

It is important to note that phrases, acronyms, digits, and special characters are not supported by Deepgram’s Keywords feature. For example, an example of something that isn’t supported by keywords is “web 3.0”. This is a phrase, so the feature will separate the terms at the space, and boost “web” and “3.0” separately.

“Web” is a very common word, it should be transcribed accurately in most cases, and boosting it could result in false positives, so it is not recommended as a keyword. “3.0” won’t work, as it includes numerals and special characters, which are not supported by this feature.

Keyword Boosting vs. Custom Model Training

If you find that you are spending a lot of time trying to configure the keyword feature in your request, it is highly likely that you need a custom-trained model that has been taught to learn and identify your key terms. Some criteria by which you can judge whether you need to train a custom model instead of using our keyword feature are the following:

• When the number of keywords you are using becomes excessive

• When there are numerous words that are outside of your existing custom model’s vocabulary

• When you are looking to emphasize terms that are not single proper nouns

Conclusion

In short, the keyword feature is meant to act more as light guidance for your existing model. If your configuration becomes too prescriptive, consider training a custom model instead.

For more information on custom model training, please see the associated product page here or contact one of our Product Consultants to learn more: https://deepgram.com/contact-us/