Simon Willison - @simonw - simonwillison.net
For more on this project: A tool to run caption extraction against online videos using Whisper and GitHub Issues/Actions
Action Transcription supports archiving and searching the transcripts of videos from multiple different video hosting platforms.
It runs on GitHub to take advantage of the free GitHub Actions code running mechanism - but importantly it does not require any use any tools aside from the user's browser, even to setup new instances of the tool.
If a video has captions, this tool can be used to retrieve and store those captions.
If a video does not have captions, the tool can extract the audio from the video and run it through Whisper - a new, state-of-the-art speech-to-text tool from OpenAI.
A demo version of this tool can be found at simonw/action-transcription-demo.
- Example issue with a VK video transcribed to English using Whisper
- Example issue that extracted YouTube auto-generated English captions
This GitHub repository acts as a "template repository" - you can create your own copy of the repository using this form.
These can be created public or private - public repos get an additional feature and are free to run, while private repos have additional cost.
If you wish to use the "Whisper" integration you will need to create an account on Replicate, then copy the API token from that account and create a new GitHub Actions secret in your repository called REPLICATE_API_TOKEN. Transcribing videos costs money - usually around $0.20 per minute of audio.
Usage of the tool is through filing GitHub Issues.
Issues must include the URL to the video you want to transcribe in the issue body.
You can tag the issue with "captions" to extract captions from the video hosting provider (which is free), or "whisper" to transcribe the audio using Whisper (which costs money).
Note that "whisper" transcriptions only work on shorter videos: up to five minutes should be OK, but longer than that is likely to fail with a timeout error.
In public repos, issue templates are provided which help further guide the user through the process. Here's a demo:
The results of the operation will be posted in a comment on the issue, and will also be written to the GitHub repository for permanent storage.
The pattern I am most excited about here is the way this shows that GitHub Issues can be used to create a hopefully not-too-intimidating interface for users, which can trigger real code to be run for free by the GitHub Actions platform.
My next step with this project will be to add a custom search engine that can be used to search the transcripts of the videos. I intend to build this using Datasette Lite.
Original development took place in the issues in the action-transcription-prototype repository.