I've been working with a client over the past month to improve Fireflies.ai's transcription accuracy for their internal meetings, specifically around their proprietary product names and technical jargon. The out-of-the-box recognition for these terms was hovering around 40%, which was causing significant friction during review.
We dedicated time to systematically train the custom vocabulary feature. The process was straightforward: we compiled a list of about 50 key terms, including product names, internal project codenames, and a few key client acronyms. We uploaded the list, provided clear pronunciations, and linked terms to relevant context.
After a week of the model processing this data across new recordings, we saw a marked improvement. In our latest batch of meeting transcripts, accuracy for those trained terms jumped to approximately 70%. This has made the search functionality far more reliable and reduced the time spent correcting crucial nouns.
A few observations from the exercise:
* The training is clearly cumulative; the more meetings Fireflies processes with the vocabulary active, the better it seems to get.
* It's important to be precise with pronunciation entries, especially for acronyms spoken as letters (e.g., "C-I-A" vs. "cia").
* Common words that are also product names (e.g., "Arc," "Portal") still require context, so you'll see some persistent errors, but the trend is positive.
For teams with a specialized lexicon, this feature is worth the investment of an hour to set up. I'm curious if others have done similar vocabulary training and what your results have been. Did you hit a ceiling, or did accuracy continue to climb over a longer period? Any pitfalls to avoid when managing the custom word list?