Request for the Ukrainian Piper High Training Checkpoint (.ckpt)
Hi Roman,
Thank you for publishing the Ukrainian Piper high voices — they sound excellent.
I’m building a personal Ukrainian voice model and have prepared about 60 minutes of clean, aligned speech data. I’d like to fine-tune a Ukrainian high Piper model with this dataset.
Would it be possible to share the original training checkpoint (.ckpt) and, if available, the training configuration used for uk_UA-mykyta-high, uk_UA-oleksa-high, or uk_UA-tetiana-high?
The published ONNX models work perfectly for inference, but they cannot be used for fine-tuning. I would use the checkpoint only for personal experimentation and local Home Assistant TTS.
Thank you!
Hi,
Thanks for reaching out! It's great to hear that the Ukrainian Piper voices are working well for you.
I've created a dedicated repository containing the original training checkpoints (.ckpt) for Mykyta, Oleksa, and Tetiana: https://huggingface.co/RomanStasyshyn/uk_UA-piper-checkpoints
Note: HF's automated scanner flags these files as "Suspicious" — this is a known false positive for PyTorch checkpoints (triggered by standard pathlib/collections imports). JFrog's scan confirms no issues, so it's safe to load with standard torch.load().
Best of luck with your personal voice model and Home Assistant TTS setup!
Best regards,
Roman
Hi Roman,
Thank you very much for creating and sharing the dedicated checkpoint repository so quickly!
I appreciate the clarification about the Hugging Face scanner warning as well.
Thanks again for making Ukrainian TTS more accessible!
Best regards,
Oleh