Pete Warden and team just published a paper on Moonshine, their speech to text model.
Key features include:
- 1.7x overall speed boost compared to Whisper
- Flexible-sized input window, allowing for more efficient processing of shorter audio clips
- Up to 5x faster performance on 10-second audio clips
- Matches or exceeds Whisper's accuracy
This is exactly right—we now live in a world in which most jobs are knowledge work, and we should look to those who are the most productive (and lazy) knowledge workers: software developers.
Yeah. I’m convinced the current model is just too confusing. But I really wish there were new interaction patterns that took advantage of low latency speech recognition...
Key features include:
- 1.7x overall speed boost compared to Whisper - Flexible-sized input window, allowing for more efficient processing of shorter audio clips - Up to 5x faster performance on 10-second audio clips - Matches or exceeds Whisper's accuracy