skypro1111/fish-speech-1.5-ukrainian

Fish Speech 1.5 Ukrainian

This is a fine-tuned version of Fish Speech 1.5 model, specifically trained for Ukrainian language.

Model Details

• Base Model: Fish Speech 1.5
• Dataset Size: 240 hours
• Number of Speakers: 55
• Training Steps: 100,000
• Training Process: Followed the official Fish Speech fine-tuning guide

Training Dataset

The model was fine-tuned on a private dataset containing:

• 240 hours of Ukrainian speech
• 55 unique speakers
• High-quality studio recordings
• Various speech patterns and intonations

Fine-tuning Process

The fine-tuning process strictly followed the official Fish Speech documentation:

1. Dataset preparation in the required format
2. Batch extraction of semantic tokens using VQGAN
3. Dataset packing into protobuf format
4. LoRA fine-tuning of the LLAMA component

License

This model is released under the same license as the original Fish Speech model - CC BY-NC-SA 4.0.

This means you:

• Must give appropriate credit
• Cannot use the model for commercial purposes
• Must distribute any modifications under the same license
• Cannot apply legal terms that legally restrict others from doing anything the license permits

Acknowledgments

• Original Fish Speech model by Fish Audio
• Training infrastructure provided by Fish Speech