itzune/maider-dataset

Maider Dataset (Synthetic)

This is a large-scale synthetic speech corpus designed for training and fine-tuning Basque Text-to-Speech (TTS) models. It consists of 99,996 audio files synthesized from the "Maider" voice model.

This dataset was generated by Itzune and serves as the primary source for training the itzune/maider-tts (Piper version) model.

Dataset Structure

Due to the large volume of data (approx. 100,000 files), the dataset is organized in the WebDataset format. The audio files are bundled into .tar shards to optimize storage, I/O performance, and streaming.

Files

• data/: Directory containing the .tar shards.
• metadata.csv: The main metadata file using | as a delimiter:
  • file_name: The name of the audio file (e.g., audio_1.wav).
  • transcription: The corresponding Basque text.

Technical Specifications

• Audio Format: WAV (PCM)
• Sample Rate: 22050 Hz
• Language: Basque (eu)
• Voice Profile: Maider (Female)
• Text Source: Mozilla Common Voice - Basque Sentence Collection
• Generation Method: Synthesized using VITS-based architecture.

Usage

from datasets import load_dataset

dataset = load_dataset("itzune/maider-dataset", streaming=True)
sample = next(iter(dataset["train"]))
print(f"Text: {sample['transcription']}")

Credits and Licensing

Source and Methodology

This is a synthetic dataset generated by Itzune. The synthesis process involved:

• Text Acquisition: Sentences were sourced from the Mozilla Common Voice project (Basque sentence collection).

• Audio Synthesis: The audio was produced using the aHoTTS synthesis tools and the pre-trained Maider (VITS) model developed by HiTZ Basque Center for Language Technology - Aholab Signal Processing Laboratory.

Acknowledgments

Mozilla Common Voice: For providing the community-driven sentence collection.

HiTZ Basque Center for Language Technology - Aholab Signal Processing Laboratory: For the underlying synthesis technology and the Maider voice model.

Project ILENIA: The original Maider voice resource was developed with funding from Project ILENIA.

License

Dataset Content (Audio & Text): Licensed under Creative Commons Attribution 4.0 International (CC BY 4.0).

Original Tools/Code: The aHoTTS tools used to generate this data are licensed under the Apache License 2.0.

Citation

If you use this dataset, please cite the original work from HiTZ/Aholab:

García, V., Hernáez, I., & Navas, E. (2022). Evaluation of Tacotron Based Synthesizers for Spanish and Basque. Applied Sciences, 12(3), 1686. https://doi.org/10.3390/app12031686