Hi @synesthesiam,
Is there anything stopping me from using a larger model with Whisper? Would it be as simple as updating the available options here, or is more work needed to enable medium (non-quantized) and large-v2?
I understand I would probably need to use a different image with CUDA installed.
The models are downloaded from Github, so I didn’t upload the larger models. I’m thinking of shifting over to HuggingFace so it would be possible to put all of the models up.
This would allow anyone to use their own fine tuned models, right? It would be super cool.
This is already super cool
There’s very little explaining this on the repo… is this a faster replacement to the usual whisper add-on? That add-on says it uses faster-whisper already, but points to a different repo.
This is the repo of the standard addon for HA. Michael recently updated the version and custom models are now supported.
Do not expect any significant change in the operating speed. In any case, large models need a GPU. And small models generate unstable recognition outcomes.