German Language Support

[ctxadm]

Hey Gents, is it possible to integrate german language pls?

there is already a german voice from Thorsten here on github

https://huggingface.co/datasets/Thorsten-Voice/TV-44kHz-Full/tree/main/TV-2021.02-Neutral

[RBEmerson970]

AFAIK, that's a mod for Kokoro itself, which means getting support from hexgrad. Look at the Kokoro Hugging Face pages, and it doesn't seem as though there's a big welcome for "foreign" (i.e. non-hexgrad) data. IMHO I suspect Kokoro has reached "what you see is what you get" status. I doubt we'll see any updates.

Glaub mir es, ich möchte gern deutsch mit FastKoko haben.

[ctxadm]

I see it differently, as there is a lot of potential in local TTS languages. I have now found the path where the Voices.pt files are stored, it may be possible to simply store new Voices.pt files there and reload them by restarting Docker ...will test this

[RBEmerson970]

Please give it a try!

[ctxadm]

did u already tried this??

[RBEmerson970]

No. I'm a writer, and I use FastKoko to read back my work as a "does it sound right" check. I've included bits of German in a couple of my stories. German read back with an English-speaking voice is...um...unfortunate. Looking into German support, it looked as though Kokoro isn't very flexible: get one language or the other, but forget mixing languages.

[semidark]

Working on it and publishing my progress here:

Thomcle/kokoro_german#1

[semidark]

Exciting News: Kokoro TTS Training Recipe Fully Open-Sourced!

We're thrilled to announce a major breakthrough: the Kokoro TTS training recipe has now been completely liberated! By pivoting to our own repository semidark/kokoro-deutsch, we've successfully open-sourced the entire training process as of yesterday.

With this recipe now publicly available, we're actively working on training German Kokoro voice packs. You can listen to very early and rough basemodel training samples from the first of two needed Training runs here.

Important The basemodel needs a finetune run before the quality gets better. We are doing this right now.

Important Note:
While the training process is mostly solved, a high-quality German TTS dataset remains a significant challenge. We're tackling this by quality-checking existing German corpora and augmenting training data with synthetic samples generated by free and open-source TTS systems.

We welcome code contributions, dataset assistance, and GPU donations to help accelerate development. If you'd like to support the project or get involved, please reach out!