I work for a news organization. Our news team needs to transcribe a lot of content from things like city hall meetings, interviews, press briefings, etc. Often the reporter isn’t present for these, but instead reviews the meeting from a video on a social media site like YouTube or Facebook. Sometimes transcripts are available for this content, and for the times when it isn’t I put together the Mistral long transcriber.
The long transcriber is a few python scripts and an html file that leverages Lunr.js to make searchable transcripts. The tool is designed to work from the command line or using an html page. It’s my first attempt at making a self-bootstrappable python project. It works … ok. Lots to do still.
I chose Mistral (Voxtral Mini Transcribe 2 really) as the transcription tool for a few reasons. First, for a news organization that has a habit of pissing off the government I didn’t want to rely on data that was hosted in the U.S. Mistral is moving in that direction. Second, it’s a multilingual transcription tool that is also really good. Third, it is pretty cheap. Fourth, while I can’t always choose open models, choosing models by providers who also release a lot of open models is next best. Mistral has a good record in that department.
The last reason is somewhat summed up by all the rest. I’m a big believer that helping to extend Europe’s digital sovereignty is the best bet for me to have my own digital soverignty. The U.S. is pretty baked into its current situation of being ruled by oligarchs like Google, Meta, Apple, etc. and while the EU has its problems they also recognize that they even have a problem and are actively working to fix it. I want to use, build with and promote tools created by people who are doing that.
</rant>
While this transcoder is good, it’s not as easy to set up and use as a SaSS product, and it relies on tools like yt-dlp to get the audio you need to transcribe. That has created friction against its use at my place of work. The good news is there are other users of this project, namely the FARMS Project.
Mistral’s transcription tools have come a long way since I started working on this project, but it seems this little guy still fills a niche, at least until AnythingLLM starts supporting their transcription services on Linux.