meetween/mumospee_librispeech

Dataset Summary

This dataset is a modified version of the LibriSpeech corpus, converted into parquet format to enhance I/O efficiency in high-performance computing environments. LibriSpeech is widely used for speech recognition research and includes over 1,000 hours of English read speech. This modification retains the original data attributes while improving data handling for distributed training on large-scale models.

---

Source Data

• Original Dataset: LibriSpeech
• License: The original LibriSpeech dataset is licensed under the Creative Commons Attribution 4.0 International License, which applies to this derived dataset.

Modifications

• Data Format: Converted to parquet files for more efficient I/O operations.
• Optimization: Designed to reduce storage and improve data loading speed in high-performance or distributed computing setups.

Dataset Structure

• File Format: Parquet files.
• Languages: English.
• Audio Sampling Rate: 16 kHz, consistent with LibriSpeech specifications.
• Speaker Representation: Multiple speakers retained from LibriSpeech’s corpus, offering diverse voice samples for training and evaluation.

mumospee_librispeech/
├── librispeech-parquet/     # audio + transcripts, one parquet per LibriSpeech subset
│   ├── dev-clean.parquet ... train-other-500.parquet   (7 subsets)
└── metadata/                # lightweight metadata index, one parquet per subset
    ├── dev-clean.parquet ... train-other-500.parquet

The metadata/ parquet files index the audio segments and contain the following columns:

column	type	description
path	string	audio file name (e.g. 2277-149896-0000.flac)
url	string	download URL of the audio parquet shard containing this segment
type	string	always audio
duration	float64	segment duration in seconds
language	string	always en
transcript	string	transcription
tag	string	always Librispeech
split	string	train / test / validation (LibriSpeech dev-* subsets map to validation)
license	string	always CC-BY-4.0

Total segments: 292,367.

Usage

This dataset is well-suited for ASR tasks, optimized for high-throughput training in distributed environments.

Attribution

This dataset is derived from the original LibriSpeech dataset, with modifications to format. Please cite the original LibriSpeech dataset in any publications or projects using this dataset.

Changelog

2026-06-30

• Converted the metadata/ index from CSV to parquet (one file per subset).
• Updated the audio parquet file column with the audio file's basename.