Understanding Keyterms vs Keywords in Deepgram
In the Deepgram ecosystem, Keyterms and Keywords help enhance the accuracy of transcriptions by focusing on terms audio models may struggle with. However, they are distinct features supported by different models and cater to differing use-cases.
Differences Between Keyterms and Keywords
Keyterms:
- Model Support: Exclusively available in the Nova-3 model.
- Multi-word Support: Capable of recognizing and accurately transcribing both individual words and multi-word phrases.
- Vocabulary Inclusion: Works with both in-vocabulary and out-of-vocabulary terms.
- Accuracy: Delivers a higher keyword recognition rate compared to Keywords.
Keywords:
- Model Support: Used with Nova-2 and older models.
- Single-word Focus: Primarily boosts individual uncommon words that aren't recognized correctly by the model.
- Out-of-vocabulary Emphasis: Keywords focuses on improving transcription accuracy for words not previously or successfully transcribed by the model.
Best Practices
When working with Keywords, adhere to the following guidelines:
- Target Uncommon Words: Boost words that the model struggles with. Avoid boosting common words or alphanumeric sequences.
- Limit Submission: Submit each word once to avoid redundancy in processing.
- Focus on Principal Words: Submit key parts of phrases rather than full phrases for more accurate emphasis.
For Keyterms, given their multi-word and broader vocabulary support, users can focus more on complete terminology that needs improved recognition, both familiar and unfamiliar to the model.
Conclusion
Understanding the nuances between these features allows developers to leverage Deepgram's transcription capabilities more effectively. With improved recognition rates and broader support, adopting Keyterms with Nova-3 equips users with advanced transcription accuracy.
For detailed guidance on utilizing these features, refer to the Deepgram Developers Documentation for Keyterms and Keywords.
References:
If issues persist or the system behavior seems inconsistent, reach out to your Deepgram support representative (if you have one) or visit our community for assistance: https://discord.gg/deepgram