Accuracy of near real-time streaming of STT

[d-shehu]

From the experimentation I've done with Whisper v3 large, accuracy is not that great when transcribing short segments even with highly articulate, high quality audio with no noise and native English speakers. It can be as low as 50-60% accurate with a latency of 1-3 secs. The latency is fine but quality is not.

Obviously accuracy improves significantly when transcribing larger segments such as a monologue or entire meeting. It's still not close to the high 90s %. With the latter it's probably in the 80s depending on the speaker and how articulate they are.

Has anyone gotten better results and if so what approach did you take? Prompts with vocab help a bit but it's very cumbersome and brittle. Thanks.