Add app model registry

[Kostis-S-Z]

What's changing

After a finetuning job is completed, the model is uploaded to Hugging Face. If the user wants to use this specific model in the demo app locally to transcribe audio, they need to add the specific model id to the UI. This PR enables addition and deletion of model ids from the dropdown menu of the models.

How to test it

Steps to test the changes:

1. Run the demo app from the demo directory: python app.py
2. Add a model id to registry
3. Check local model_registry.json is properly updated
4. Remove a model id from registry
5. Check local model_registry.json is properly updated

[daavoo]

Before doing a review of the code, I am not sure if I understand the high level value of adding this logic.

Why don't we just add a text input so user can copy-paste the model id after finetuning?

[Kostis-S-Z]

Why don't we just add a text input so user can copy-paste the model id after finetuning?

I was thinking that a local register would make it easier for the user to use the transcription app so that they dont have to go to HF and copy/paste the mode id every time they want to use the app. I am also working on a PR for this issue, which basically lets you select two models to compare the output side-by-side. Just thought that a dropdown of the original whisper models + your finetuned versions, is more UX friendly

[daavoo]

I was thinking that a local register would make it easier for the user to use the transcription app so that they dont have to go to HF and copy/paste the mode id every time they want to use the app. I am also working on a PR for this issue, which basically lets you select two models to compare the output side-by-side. Just thought that a dropdown of the original whisper models + your finetuned versions, is more UX friendly

Since this is user-based and we double check against the hf profile, could we just list the speech-to-text models from the user profile and include them in the dropdown?

[Kostis-S-Z]

Since this is user-based and we double check against the hf profile, could we just list the speech-to-text models from the user profile and include them in the dropdown?

You mean something like this?

api = HfApi()
api.list_models(filter="speech-to-text")

[Kostis-S-Z]

I am currently having an issue with properly assigning tags in the model card, but I can implement your suggestion. Could we merge this PR however, and I implement your suggestion after I fix #33 ?

[daavoo]

I don't have a strong opinion against this but it feels like we keep adding code that makes things a little more complex without much value added for the user (copy-paste from HF url vs using the gradio UI to register a new model)

[Kostis-S-Z]

I don't have a strong opinion against this but it feels like we keep adding code that makes things a little more complex without much value added for the user (copy-paste from HF url vs using the gradio UI to register a new model)

Its a fair point. So would you suggest we just ask the user to copy paste the url in the script here?

For a bit more context, I would like to make this repo a bit more friendly to non-devs than maybe our standard audience. The Common Voice community is quite big and I think a big part of it might not be comfortable with code. Thus, I would like that the apps can stand alone, and you can get value out of them without touching any code. Thats why I am focusing so much on UI/UX.

[daavoo]

Its a fair point. So would you suggest we just ask the user to copy paste the url in the script here?

Could you just add an input text box for users to paste the url?

[Kostis-S-Z]

Its a fair point. So would you suggest we just ask the user to copy paste the url in the script here?

Could you just add an input text box for users to paste the url?

Something like this?

[Kostis-S-Z]

Well, okay note that after the #42 PR, it would look like this

[Kostis-S-Z]

Closing because of merging the feedback at #42