# LibriSpeech ASR corpus

## Facts

- Provider: Vassil Panayotov, Guoguo Chen, Daniel Povey, Sanjeev Khudanpur
- License: CC-BY-4.0 (https://www.openslr.org/12/)
- Size: 262 files
- Format: Parquet (in sampled files)
- Stored at: Hugging Face, https://huggingface.co/datasets/openslr/librispeech_asr
- Access: public, readable without an account
- Category: Speech & Audio
- Source: https://www.danielpovey.com/files/2015_icassp_librispeech.pdf
- Page: https://scigantic.com/archive/48a232cf-2ce4-5a8e-b44a-2666982a1bbf
- Keywords: Speech & Audio, speech recognition, benchmark, Parquet, Hugging Face

## About

LibriSpeech is a corpus of approximately 1000 hours of 16 kHz read English speech prepared by Vassil Panayotov with Daniel Povey, derived from LibriVox audiobooks and carefully segmented and aligned to the book text. It is the standard benchmark for automatic speech recognition (ASR): word error rate on test-clean and test-other is the most widely reported English ASR number. Speakers are split into 'clean' and 'other' pools by the WER of a reference model, giving train-clean-100, train-clean-360, train-other-500, dev and test splits. This Hugging Face repository, maintained under the openslr organisation, holds the corpus as Parquet shards in three configurations (all, clean, other), each with one folder per split (for example all/train.other.500 holds 64 Parquet files, about 31 GB; measured 2026-11-14). The Hub reports about 218 GB for the repo. Each row carries the FLAC audio, transcript, speaker id, chapter id and utterance id; open with datasets.load_dataset('openslr/librispeech_asr', 'clean') or read the shards with pyarrow or pandas. Also used for speaker identification and self-supervised speech pretraining (wav2vec 2.0, HuBERT). Released by OpenSLR (SLR12) under CC BY 4.0; cite Panayotov et al., ICASSP 2015.

## Load the data

```python
from datasets import load_dataset
ds = load_dataset("openslr/librispeech_asr")
print(ds)
```

## Data structure

File types in sampled files: Parquet

Example file: `all/test.clean/0000.parquet`

## Open it

- In a hosted notebook with the data mounted: https://scigantic.com/archive/48a232cf-2ce4-5a8e-b44a-2666982a1bbf
- Full record as JSON (schema card included): https://api.scigantic.com/api/archives/48a232cf-2ce4-5a8e-b44a-2666982a1bbf
- Over MCP: `get_schema_card` with id `48a232cf-2ce4-5a8e-b44a-2666982a1bbf` at https://api.scigantic.com/mcp

## Related topics

- [Parquet datasets](https://scigantic.com/topics/parquet)
- [Benchmark datasets](https://scigantic.com/topics/benchmark)
- [Speech & Audio datasets](https://scigantic.com/topics/speech-and-audio)
- [Speech recognition datasets](https://scigantic.com/topics/speech-recognition)
