TranscriptOMatic - Read This First
Documentation of TranscriptOMatic. A Open Source tool chain for 'good enough' TTRPG session transcriptions on a Raspberry Pi 500.
DISCLAIMER: Frozen WIP
This project has reached a status that most likely won't see any changes in the future. A Raspberry Pi 500 with just 8GB RAM is too limited to be a really good platform for a low-cost home-grown transcription tool.
Yet, TranscriptOMatic is not dead and I will give it another go on different hardware. This document will continue to exist for everybody who like to have a starting point for their own transcription projects.
Mind the constraints:
TranscriptOMatic in its current state, running on a Raspberry Pi 500, works only
- using the whisper.cpp 'tiny.en' model
- for English language meetings
- okay'ish at best
- and by creating a recording
Do NOT use the --de or --auto options.
I kept these in the documentation to show the concept how TranscriptOMatic should work, not as proof as it actually works on a Raspberry Pi 500.
Introduction
I've started this project after fiddling around with Notions "Meeting Notes" feature and running the idea of using a Speech to Text tool for our session transcripts by the members of one of my TTRPG groups.
Consensus was that a tool like this would be nice and helpful to have good transcripts and a helpful session summary without taking time out of player's lives to compile concise session notes. For me personally, as somebody with an auditory processing disorder, having a live transcript would also be incredibly helpful to be better able to understand what has been said.
Most group members weren't concerned to use a third-party service, since we wouldn't talk about 'real' or personal stuff, but only a fictional story, concerning fictional characters.
Yet, this solution wouldn't have been ideal and there are real concerns with feeding deeply personal and biometrical information, like a person's voice, to a commercial language model of a company that is based in the USA1.
That's why I started fiddling around with open-source tools and language models that would run strictly local. With no data transferred to big-tech companies, and this way also not providing information to train commercial models.
My hope is, to create a solution by cobbling together open-source tools and some shell scripts. Ideally, this solution will be easily replicable by everybody who knows their way around a UNIX/Linux command line and has a device available with enough computing-power, RAM and/or GPU.
What TranscriptOMatic most likely won't accomplish anytime soon, is to discern who is speaking.
1) And yes, I'm aware of the irony that we already transfer our voices to a US-based service by using Discord voice chat. Doesn't mean one should make it worse.
WTF a Raspberry Pi 500?
When I began working on this idea, it was the hardware I had at hand.