This is a Proof Of Concept project to implement a version of Ian Macartney's excellent Work Stealing strategy for Whisper based transcription.
This project heavily relies upon ffmpeg, whisper-standalone-win and many other excellent technologies, so big thanks to them for doing the hard work.
Demo video and further technical details can be found on my blog post: https://mikecann.co.uk/
bun install
Make a .env.local file in the root of the project, checkout example.env.local for the kind of env vars you will need.
In one terminal bun dev:ts to dev types (I like to do this in a separate terminal as Webstorm provides clickable links to the errors if I do)
In another terminal bun dev to start Vite and Convex
The worker can either be tested in the docker container with bun docker:build-and-run or outside the container with bun start:worker to do that you will need to provide whisper on the PATH by downloading a binary from: https://github.com/Purfview/whisper-standalone-win and stick it on you PATH.
Use 'dive' to explore a Docker image contents: https://github.com/wagoodman/dive
I was testing this on Windows, no idea if it works on other platforms.
- The transcription result might be larger than the 1mb row limit in the database (https://docs.convex.dev/production/state/limits) so we might have to store it as a file instead
- There is virtually no error handling around fly machines starting up etc, so this is definately required to make sure we can recover from errors
- There is not much in the way of testing at all at the moment, so this makes me a bit nervous about the fragile code
- In addition to the above the code is rather messy and really could do with tidying up
- There is no authorization or authentication in place so this is obviously not great
- There is no limits on the number of workers that can be started, so this could be a problem
- The UI is really very basic and could do with a lot of work and is more of an experiment than anything else for me to tinker with Shadcn and Tailwind