Integrate new Kroko models?

[iGerman00]

Handy is probably one of the best apps I've used in a long time, for a while nothing has "just worked" like that in this space. I stumbled upon https://kroko.ai recently and thought to bring it up here for you.

They have very small (~70MB) and quick models and they are pretty good at ASR. Their inference engine is Apache 2.0 licensed, and their pre-trained weights are available under CC-BY-SA.

Some pitfalls, potentially:

• CC-BY-SA is probably also NC/non-commercial; it's implied in their marketing but they don't mention BY-SA-NC
• Since you take donations on https://handy.computer, that might interfere if they're stringent about it. Someone from @kroko-ai (@maxnzv, @todyal) should clarify.
• The models are not automatically multilingual, but it can definitely be worth it for really low-powered devices, or just not interfering as much with other programs.

[joazoa]

I'm with Kroko

The 70MB models are our older models. you are referring to, they were for non commercial use.

The models released 5 days ago here: https://huggingface.co/Banafo/Kroko-ASR/tree/main (those ending on .data) are CC-BY-SA, it is ok for commercial use, only requires attribution. (they are better than the older generation).

You can choose to also support the commercial models or strip that part out completely. (we have a compile flag for that).

Feel free to reach out to us to discuss this more.

[cjpais]

Happy to include it in transcribe-rs at the very least, especially if someone wants to open a PR for it :) otherwise I'll try and get around to it when I can.

@joazoa Happy to use the non-commercial here and attribute in the UI of the app (about section) and README here if and when support comes in.

Overall I'm not 100% sure I'll be adding more support for more models because it may get confusing for users who don't have a strong grasp on the ecosystem. Maybe in advanced or something. But I will rethink how model support should go and onboarding at some point