Large neural network models in the archive?

[Petter Reinholdtsen <pere@hungry.com>]

Hi.

I am considering uploading the OpenAI Whisper neural network models to
non-free, to allow well working speech to text to work out of the box
using 'apt install openai-whisper'.  The models are quite large, 2.9 GiB
for the large model, 1.5 GiB for the medium one and 462 MiB for the
small one.  What is the view of the ftpmasters on providing such source
and binary packages in Debian.  I doubt the models will change often,
but do not know for sure.

<URL: https://salsa.debian.org/deeplearning-team/openai-whisper-model.git >
is the source of the package.  I did not commit the models to git (yet).
Might not really need or want to, suspect it is enough that these models
are in the archive and snapshot, not convinced we need a copy on salsa
too.

The complete set of packages needed to get openai-whisper working is
tiktoken, triton, openai-whisper and openai-whisper-model-<size>.  All
of them are working on my test machine.

-- 
Happy hacking
Petter Reinholdtsen