metadata
license: cc-by-4.0
language:
- en
Dataset Summary
This dataset is a derived version of the LibriTTS corpus, converted into larger parquet files for optimized I/O performance on high-performance computing clusters. The dataset maintains the high-quality, multi-speaker, text-to-speech (TTS) alignment of LibriTTS, with over 585 hours of English audiobook recordings and corresponding transcriptions. This format is ideal for large-scale training in speech synthesis and TTS tasks.
Source Data
- Original Dataset: LibriTTS
- License: The original LibriTTS dataset is licensed under the Creative Commons Attribution 4.0 International License. This derived dataset retains the same license.
Modifications
- Data Format: The data has been restructured into larger parquet files to enhance I/O efficiency, reducing load times for distributed training environments.
- Storage Optimization: This derived dataset improves upon the storage requirements and retrieval efficiency, leveraging the parquet format's compression capabilities.
Dataset Structure
- File Format: Parquet files.
- Sampling Rate: 24 kHz (same as LibriTTS).
- Speaker Details: Over 2,400 unique speakers with balanced representation of male and female voices, retained from LibriTTS.
Attribution
This dataset is based on work by LibriTTS, with modifications for I/O efficiency by converting to parquet file format. Please cite the original LibriTTS dataset in any publications or projects.