# LoquaciousSet (LargeScaleASR): 25,000 hours of English speech

## Facts

- Provider: Titouan Parcollet, Yuan Tseng, Shucong Zhang, Rogier van Dalen (SpeechBrain, Samsung AI Center Cambridge)
- License: CC-BY-4.0 (https://huggingface.co/datasets/speechbrain/LoquaciousSet)
- Size: 9,756 files
- Format: Parquet (in sampled files)
- Stored at: Hugging Face, https://huggingface.co/datasets/speechbrain/LoquaciousSet
- Access: public, readable without an account
- Category: Speech & Audio
- Source: https://arxiv.org/abs/2505.21578
- Page: https://scigantic.com/archive/43ea6f4a-fdba-5e2c-beb2-a5017bea940a
- Keywords: Speech & Audio, speech recognition, Parquet, Hugging Face

## About

LoquaciousSet (also published as LargeScaleASR) is a 25,000-hour corpus of transcribed, heterogeneous English speech for research and commercial speech recognition. It mixes read and spontaneous, clean and noisy speech drawn from five openly licensed corpora: VoxPopuli (CC0), LibriHeavy (CC BY 4.0), LibriSpeech dev-other and test-other (CC BY 4.0), YODAS (CC BY 3.0), People's Speech (CC BY 4.0) and Common Voice 18.0 (CC0), with the mixing and curation described in the accompanying paper. Six subsets are provided: large (25,000 hours), medium (2,500), small (250), clean (13,000 hours, excluding YODAS and People's Speech), dev (15 hours) and test (21 hours). The repository holds those as folders of Parquet shards (the clean folder lists at least 1,000 shard files), plus per-split CSV manifests, a lexicon, a Sequitur G2P model and pruned and unpruned 3-gram and 4-gram ARPA language models (measured 2026-11-14; the Hub reports about 3.8 TB). Load with datasets.load_dataset('speechbrain/LoquaciousSet', 'small') and decode the 'wav' bytes with torchaudio; a SpeechBrain conformer recipe is linked. Every component is open; attribute each source corpus as listed on the card.

## Load the data

```python
from datasets import load_dataset
ds = load_dataset("speechbrain/LoquaciousSet")
print(ds)
```

## Data structure

File types in sampled files: Parquet

Example file: `clean/train-00000-of-02615.parquet`

## Open it

- In a hosted notebook with the data mounted: https://scigantic.com/archive/43ea6f4a-fdba-5e2c-beb2-a5017bea940a
- Full record as JSON (schema card included): https://api.scigantic.com/api/archives/43ea6f4a-fdba-5e2c-beb2-a5017bea940a
- Over MCP: `get_schema_card` with id `43ea6f4a-fdba-5e2c-beb2-a5017bea940a` at https://api.scigantic.com/mcp

## Related topics

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