itzune/maider-tts

Basque TTS: Maider (Piper Version)

This repository contains a Piper compatible version of the Maider Basque text-to-speech model. The original model was developed by HiTZ Basque Center for Language Technology - Aholab Signal Processing Laboratory (University of the Basque Country UPV/EHU).

This version has been exported/trained specifically for use with the Piper TTS engine, a fast, local neural text-to-speech engine.

Model Details

• Language: Basque (eu)
• Speaker: Maider (Female)
• Architecture: VITS (Optimized for Piper)
• Original Credits: HiTZ Center / Aholab (Project ILENIA)
• Format: Piper (.onnx and .onnx.json config)

Training Details

• Dataset: itzune/maider-dataset
• Data Volume: 99,996 high-quality audio samples (~100k files).
• Architecture: VITS
• Training Engine: Piper (PyTorch Lightning)
• Iterations: 22 epochs (258,750 steps)
• Sample Rate: 22050 Hz
• Phonemization: espeak-ng (Basque)

Data Source & Dataset Integration

This model has been fine-tuned/trained using the maider_dataset, a large-scale Basque speech corpus specifically curated for high-fidelity synthesis.

• Link to Dataset: Maider Dataset on Hugging Face
• Dataset structure: The data was processed using WebDataset (sharded .tar files) to handle the 100k samples efficiently after the Piper training process.
• Content: Each audio file is paired with its corresponding Basque transcription in a metadata.csv file, ensuring precise alignment during the 22 epochs of training.

Files Included

• eu-maider-medium.onnx: The exported model for fast inference.
• eu-maider-medium.onnx.json: The configuration file (includes phoneme map and synthesis settings).
• epoch=22-step=258750.ckpt: The PyTorch Lightning checkpoint from the 22nd iteration (useful for further training/fine-tuning).

Usage

Using Piper CLI

You can run the model locally using the Piper binary:

echo "Kaixo, hau Maider da, Piper motorra erabiliz euskaraz hitz egiten." | \
  ./piper --model eu-maider-medium.onnx --output_file output.wav

Python API

from piper.voice import PiperVoice

voice = PiperVoice.load("eu-maider-medium.onnx", "eu-maider-medium.onnx.json")
with open("output.wav", "wb") as f:
    voice.synthesize_wav("Gaur egun eguzkitsua dugu.", f)

Original Model & Data Source

The base model belongs to the Aholab TTS collection. All voices in this collection are based on the VITS architecture proposed by Kim et al. (2021).

Maider & Antton: Developed by HiTZ with funding from Project ILENIA.

License: Public Creative Commons Attribution 4.0 (for the voice resource) and Apache License 2.0 (for the code/model).

Authors & Credits

The original Maider model was created by: HiTZ Basque Center for Language Technology - Aholab Signal Processing Laboratory, University of the Basque Country EHU.