Voices

[zha0jiapeng]

I would like to ask where to get these npy files. I went to the source project and found that they were all bin files. I don’t know how to convert them, and when I use the hexgrad/Kokoro-82M-v1.1-zh model, I still get an error.

[Lyrcaxis]

Hi. The voices were converted from the KokoroSharpBinaries repo like so:
https://github.com/Lyrcaxis/KokoroSharpBinaries/blob/main/convert_kokoro_voices.py

Then an automated release was created from the github workflows:
https://github.com/Lyrcaxis/KokoroSharpBinaries/blob/main/.github/workflows/convert-voices.yml
https://github.com/Lyrcaxis/KokoroSharpBinaries/blob/main/.github/workflows/release.yml

---

To run the zh model you just gotta ensure the tensor names match what KokoroSharp expects when exporting to ONNX.

I managed to make this work locally but wanted to set up something automated, which had bigger complexity, so it got stalled 😅

If you (or someone else) are up to contributing, it would be awesome to have the model & voice files uploaded in the binaries repo via an automation from that repo, then the remaining changes from KokoroSharp side would be simple (e.g.: adding KModel.ZH option, etc).

[zha0jiapeng]

Sorry, I'm still new to models. I only found kokoro-v1_1-zh.pth. How can I convert it into onnx to use in kokoroCharp?

[Lyrcaxis]

@zha0jiapeng you'd have to clone the kokoro repo from GitHub (or install the python package), and run the export.py script, after doing the minor tweaks shown in the image.

I could likely find the model and upload it temporarily sometime today, if you think converting it yourself is too much of a drag.