This is a Welsh (Cymraeg) neural text-to-speech model trained using Piper, a fast, local neural TTS system optimized for Raspberry Pi and other low-end devices.

Developed by: Uned Technolegau Iaith (Language Technologies Unit), Bangor University Model type: Neural TTS (VITS-based architecture) Language: Welsh (cy_GB) License: CC0-1.0 Format: ONNX

Model Details

Architecture: Based on Piper's VITS (Variational Inference with adversarial learning for end-to-end Text-to-Speech)
Speakers: Multi-speaker model with 3 speaker variants
Quality: Medium quality (suitable for screen readers and assistive technology)
Model Size: Approximately 77 MB
Inference Speed: Optimized for real-time synthesis on CPU
Sample Rate: 22050 Hz
Training Framework: Piper training pipeline

Training Data

This model was trained on the bu-tts-cy-en dataset (Bangor University Text to Speech Welsh-English dataset).

Dataset characteristics:

Size: 10,000-100,000 samples
Languages: Welsh and English (bilingual dataset)
License: CC0 1.0 (Public Domain)
Content: Audio recordings with corresponding text transcriptions
Source: Language Technologies Unit, Bangor University

Training data limitations:

Dataset consists of freely available recordings (public domain audiobooks and research-quality recordings)
Coverage is not comprehensive across all Welsh vocabulary and contexts
Some pronunciation patterns may be influenced by the limited speaker diversity in the training data
Quality improvements would be possible with larger, more diverse, professionally-recorded datasets

Intended Use

Primary use cases:

Screen readers and assistive technology (particularly NVDA integration)
Accessibility tools for Welsh speakers with visual impairments
Welsh language learning applications
Local, offline Welsh TTS applications
Research in Welsh speech synthesis

Supported platforms:

Compatible with Piper TTS runtime
Works with Sonata TTS engine
ONNX Runtime on x86/x64 architectures
Raspberry Pi and other resource-constrained devices

Usage

With Piper

bash

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

# Run synthesis
echo "Bore da, sut wyt ti?" | piper \
  --model cy_GB-bu_tts.onnx \
  --output_file output.wav

With NVDA Screen Reader

Install the techiaith Welsh Neural Voices addon for NVDA:

Download the addon from the releases page
Install and restart NVDA
Voices will download automatically on first run (77 MB)
Select "Uned Technolegau Iaith - Welsh Neural Voices" in NVDA's speech settings

With Python (ONNX Runtime)

python

import onnxruntime as ort
import numpy as np
import json
import wave

# Load model
session = ort.InferenceSession("cy_GB-bu_tts.onnx")

# Load config
with open("cy_GB-bu_tts.onnx.json") as f:
    config = json.load(f)

# For complete implementation, refer to:
# https://github.com/rhasspy/piper/blob/master/src/python_run/piper/voice.py

With Sonata Engine

python

from sonata import tts_engine

engine = tts_engine.TTSEngine()
engine.load_voice("cy_GB-bu_tts.onnx")

# Synthesize speech
audio = engine.synthesize("Bore da!")
engine.save_audio(audio, "output.wav")

Sample Audio

Listen to voice samples at: Piper Welsh samples

Limitations

Pronunciation: May exhibit incorrect or unusual pronunciation for some words, particularly:
- Technical terms and neologisms
- Place names not represented in training data
- Words with ambiguous pronunciation rules
Audio Quality: Medium quality - suitable for assistive technology but not studio-grade
Domain Coverage: Best performance on general conversational text; may struggle with specialized domains
Expressivity: Limited emotional range (neutral/informative tone)
Platform: Optimized for CPU inference on x86/x64; ARM64 Windows not supported
Language Mixing: While trained on bilingual data, best results when using pure Welsh text

Performance

Real-time Factor: < 1.0 on modern CPUs (faster than real-time synthesis)
Latency: Low latency suitable for interactive applications
Memory Usage: ~100 MB RAM during inference
Supported Platforms: Windows 10/11 (x86/x64), Linux (x86/x64), Raspberry Pi

Model Files

This repository contains:

cy_GB-bu_tts.onnx - The neural TTS model in ONNX format
cy_GB-bu_tts.onnx.json - Model configuration file (phoneme mapping, sample rate, etc.)

Citation

If you use this model, please cite:

bibtex

@misc{cy_GB_bu_tts_2025,
  author = {{Language Technologies Unit, Bangor University}},
  title = {cy\_GB-bu\_tts: Welsh Neural Text-to-Speech Model},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/techiaith/cy_GB-bu_tts}}
}

@dataset{bu_tts_cy_en_2025,
  author = {{Language Technologies Unit, Bangor University}},
  title = {Bangor University Text to Speech Welsh-English Dataset},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/techiaith/bu-tts-cy-en}}
}

@misc{piper_tts,
  author = {{Rhasspy Community}},
  title = {Piper: A fast, local neural text to speech system},
  year = {2023},
  publisher = {GitHub},
  howpublished = {\url{https://github.com/rhasspy/piper}},
  note = {Now maintained at \url{https://github.com/OHF-Voice/piper1-gpl}}
}

Acknowledgments

This work builds upon contributions from the wider open-source TTS community:

Piper TTS and the Rhasspy community for developing the training framework and TTS architecture that makes high-quality, local neural TTS accessible
Musharraf Omer for creating Sonata TTS engine and the Sonata-NVDA addon, which enables seamless integration with screen readers
Contributors to the Welsh language TTS training data
The broader open-source speech synthesis community for advancing accessible voice technology

License

This model is released under CC0-1.0 (Public Domain). You are free to use, modify, and distribute this model for any purpose without restriction.

The training code (Piper) is licensed under MIT License.

Contact & Support

Organization: Uned Technolegau Iaith / Language Technologies Unit, Bangor University Issues: Report issues at GitHub Issues Project Page: NVDA Welsh Neural Voices

Version History

2025.11.0 (Beta): Initial public release with 3 speaker variants, medium quality

Related Resources

NVDA Welsh Neural Voices Addon - Screen reader integration
Piper TTS - Training and inference framework
Sonata Engine - Cross-platform TTS engine
Training Dataset - Welsh-English TTS corpus

This model was developed to support Welsh language accessibility and to preserve and promote the Welsh language through modern speech technology.