MODEL_CARD.md

cy_GB-bu_tts - Welsh TTS Model

This is a Welsh (Cymraeg) text-to-speech model trained using Piper.

Model Details

• Language: Welsh (cy) - Monolingual
• Sample Rate: 22,050 Hz
• Quality: Medium
• Framework: Piper/VITS
• Model Type: Single-speaker TTS
• Training Dataset: techiaith/bu-tts-cy-en (Welsh samples only)

This model was trained exclusively on Welsh language data from the techiaith/bu-tts-cy-en dataset.

Usage

With Piper

Install Piper and use the model:

# Download the model files
wget https://huggingface.co/techiaith/cy_GB-bu_tts/resolve/main/cy_GB-bu_tts-medium.onnx
wget https://huggingface.co/techiaith/cy_GB-bu_tts/resolve/main/cy_GB-bu_tts-medium.onnx.json

# Generate speech
echo "Helo, sut ydych chi heddiw?" | piper --model cy_GB-bu_tts-medium.onnx --output_file output.wav

Sample Sentences

Try these example sentences in Welsh:

• "Bore da, sut ydych chi?"
• "Diolch yn fawr am eich help."
• "Mae'r tywydd yn hyfryd heddiw."
• "Rwy'n dysgu siarad Cymraeg."
• "Croeso i Gymru!"

Training Information

• Base Model: Fine-tuned from existing Piper checkpoint
• Training Configuration:
  • Batch size: 32
  • Validation split: 10%
  • Max epochs: 200
  • Precision: FP16 (mixed precision)
  • Checkpoint frequency: Every 5 epochs
  • Gradient clipping: 1.0
  • Warmup epochs: 5
  • Language filter: Welsh (cy) only

License

MIT License - See the training repository for details.

Citation

If you use this model, please cite:

@misc{cy_GB_bu_tts,
  title={Welsh Text-to-Speech Model},
  author={Uned Technolegau Iaith, Prifysgol Bangor},
  year={2025},
  publisher={HuggingFace},
  howpublished={\url{https://huggingface.co/techiaith/cy_GB-bu_tts}}
}

Acknowledgments

• Built using Piper by Rhasspy
• Training data from the techiaith project