While visiting a convention during Easter, it occurred to me that it would be great if I could have a digital Dictaphone with transcribing capabilities, providing me with texts to cut-n-paste into stuff I need to write. The background is that long drives often bring up the urge to write on texts I am working on, which of course is out of the question while driving. With the release of OpenAI Whisper, this seem to be within reach with Free Software, so I decided to give it a go. OpenAI Whisper is a Linux based neural network system to read in audio files and provide text representation of the speech in that audio recording. It handle multiple languages and according to its creators even can translate into a different language than the spoken one. I have not tested the latter feature. It can either use the CPU or a GPU with CUDA support. As far as I can tell, CUDA in practice limit that feature to NVidia graphics cards. I have few of those, as they do not work great with free software drivers, and have not tested the GPU option. While looking into the matter, I did discover some work to provide CUDA support on non-NVidia GPUs, and some work with the library used by Whisper to port it to other GPUs, but have not spent much time looking into GPU support yet. I've so far used an old X220 laptop as my test machine, and only transcribed using its CPU.
As it from a privacy standpoint is unthinkable to use computers under control of someone else (aka a "cloud" service) to transcribe ones thoughts and personal notes, I want to run the transcribing system locally on my own computers. The only sensible approach to me is to make the effort I put into this available for any Linux user and to upload the needed packages into Debian. Looking at Debian Bookworm, I discovered that only three packages were missing, tiktoken, triton, and openai-whisper. For a while I also believed ffmpeg-python was needed, but as its upstream seem to have vanished I found it safer to rewrite whisper to stop depending on in than to introduce ffmpeg-python into Debian. I decided to place these packages under the umbrella of the Debian Deep Learning Team, which seem like the best team to look after such packages. Discussing the topic within the group also made me aware that the triton package was already a future dependency of newer versions of the torch package being planned, and would be needed after Bookworm is released.
All required code packages have been now waiting in the Debian NEW queue since Wednesday, heading for Debian Experimental until Bookworm is released. An unsolved issue is how to handle the neural network models used by Whisper. The default behaviour of Whisper is to require Internet connectivity and download the model requested to ~/.cache/whisper/ on first invocation. This obviously would fail the deserted island test of free software as the Debian packages would be unusable for someone stranded with only the Debian archive and solar powered computer on a deserted island.
Because of this, I would love to include the models in the Debian mirror system. This is problematic, as the models are very large files, which would put a heavy strain on the Debian mirror infrastructure around the globe. The strain would be even higher if the models change often, which luckily as far as I can tell they do not. The small model, which according to its creator is most useful for English and in my experience is not doing a great job there either, is 462 MiB (deb is 414 MiB). The medium model, which to me seem to handle English speech fairly well is 1.5 GiB (deb is 1.3 GiB) and the large model is 2.9 GiB (deb is 2.6 GiB). I would assume everyone with enough resources would prefer to use the large model for highest quality. I believe the models themselves would have to go into the non-free part of the Debian archive, as they are not really including any useful source code for updating the models. The "source", aka the model training set, according to the creators consist of "680,000 hours of multilingual and multitask supervised data collected from the web", which to me reads material with both unknown copyright terms, unavailable to the general public. In other words, the source is not available according to the Debian Free Software Guidelines and the model should be considered non-free.
I asked the Debian FTP masters for advice regarding uploading a model package on their IRC channel, and based on the feedback there it is still unclear to me if such package would be accepted into the archive. In any case I wrote build rules for a OpenAI Whisper model package and modified the Whisper code base to prefer shared files under /usr/ and /var/ over user specific files in ~/.cache/whisper/ to be able to use these model packages, to prepare for such possibility. One solution might be to include only one of the models (small or medium, I guess) in the Debian archive, and ask people to download the others from the Internet. Not quite sure what to do here, and advice is most welcome (use the debian-ai mailing list).
To make it easier to test the new packages while I wait for them to clear the NEW queue, I created an APT source targeting bookworm. I selected Bookworm instead of Bullseye, even though I know the latter would reach more users, is that some of the required dependencies are missing from Bullseye and I during this phase of testing did not want to backport a lot of packages just to get up and running.
Here is a recipe to run as user root if you want to test OpenAI Whisper using Debian packages on your Debian Bookworm installation, first adding the APT repository GPG key to the list of trusted keys, then setting up the APT repository and finally installing the packages and one of the models:
curl https://geekbay.nuug.no/~pere/openai-whisper/D78F5C4796F353D211B119E28200D9B589641240.asc \ -o /etc/apt/trusted.gpg.d/pere-whisper.asc mkdir -p /etc/apt/sources.list.d cat > /etc/apt/sources.list.d/pere-whisper.list <<EOF deb https://geekbay.nuug.no/~pere/openai-whisper/ bookworm main deb-src https://geekbay.nuug.no/~pere/openai-whisper/ bookworm main EOF apt update apt install openai-whisper
The package work for me, but have not yet been tested on any other computer than my own. With it, I have been able to (badly) transcribe a 2 minute 40 second Norwegian audio clip to test using the small model. This took 11 minutes and around 2.2 GiB of RAM. Transcribing the same file with the medium model gave a accurate text in 77 minutes using around 5.2 GiB of RAM. My test machine had too little memory to test the large model, which I believe require 11 GiB of RAM. In short, this now work for me using Debian packages, and I hope it will for you and everyone else once the packages enter Debian.
Now I can start on the audio recording part of this project.
As usual, if you use Bitcoin and want to show your support of my activities, please send Bitcoin donations to my address 15oWEoG9dUPovwmUL9KWAnYRtNJEkP1u1b.