hexgrad/Kokoro-82M V1

[n00b001]

V1 of Kokoro has been released

Does this mean the documentation on how to pull the model and related files needs to be updated?

As a side note, instead of updating the docs each time, would it be easier if speaches-ai downloaded the relevant files itself?

[fedirz]

I will make the necessary changes to support hexgrad/Kokoro-82M V1 this week.

Does this mean the documentation on how to pull the model and related files needs to be updated?

It's likely, although I'm thinking of how to make the setup experience pain-free, just like with Whisper models...

As a side note, instead of updating the docs each time, would it be easier if speaches-ai downloaded the relevant files itself?

The setup for Kokoro is not trivial and requires downloading a mix of files from both HuggingFace and GitHub. I want to be mindful of downloading random binary files from multiple data sources. Piper also has its complexities where dozens of models live in the same HF repo.

One approach I'm considering is creating an HF repo for each Kokoro and Piper model, including all the relevant files (a 1:1 relationship between a repo and a model).
Pros:

• There will be a defined HF repo structure, making it easy for others to replicate it and add support for custom models.
  Cons:
• Maintaining all those model repos. I'll need to develop and maintain a naming scheme...