How domain context boosts post-call transcription quality (Full Transcript)
[00:00:00] Speaker 1: Awesome. I heard a few things around like accuracy, latency, code switching, speaker diarization. I guess probably all of them matter to some degree, but like maybe talk to us about like what are the ones that matter most and maybe what are some of the like things that you've done to make sure that like reliably your app actually is able to use voice in production for those particular features.
[00:00:19] Speaker 2: So for us actually latency is not that big of an issue because like we first upload the file kind of and we do all the summary after after the call or after the recording is being uploaded. Depends if it's telehealth or in person. We do some extra steps though, so kind of to double check that everything is correct. So for example here we could see on a live demo how we like improve the vocabulary because it has some context. We use those key terms as well, but what we do also is like we give the context of the visit. We mostly work with women's health, so we have certain words or diagnosis that of course like repeat themselves. So we kind of plug this into a prompt. So LLM then kind of goes through the transcript and tries to improve it if there's like anything out of the context. That's I would say it's like a biggest improvement for us. Those like improving models is really changing like there's less and less changes actually that you notice compared to like initial transcription.
- Latency may be less important when transcription/summarization is performed post-call rather than live.
- Accuracy can be improved by providing visit context and a curated domain vocabulary to guide corrections.
- LLM-based post-processing can detect and fix out-of-context transcription errors.
- Adding validation/double-check steps increases reliability for production use.
- As base ASR models improve, the amount of necessary post-correction tends to decrease.