# The People's Speech: 30,000-hour English speech recognition corpus

## Facts

- Provider: MLCommons (Daniel Galvez et al.)
- License: CC-BY-SA-4.0 (https://huggingface.co/datasets/MLCommons/peoples_speech)
- Size: 4,498 files
- Format: Parquet (in sampled files)
- Stored at: Hugging Face, https://huggingface.co/datasets/MLCommons/peoples_speech
- Access: public, readable without an account
- Category: Speech & Audio
- Source: https://arxiv.org/abs/2111.09344
- Page: https://scigantic.com/archive/7894b47e-8924-5292-b901-84da6b644482
- Keywords: Speech & Audio, speech recognition, Parquet, Hugging Face

## About

The People's Speech is among the largest English speech recognition corpora, with more than 30,000 hours of transcribed English speech from a diverse set of speakers, collected by MLCommons from openly licensed audio on the Internet Archive, including legal and government proceedings, oral histories and speeches. It was built to give speech-to-text research and industry a corpus large enough to train modern ASR systems under a permissive licence. Data are split by licence and by transcript quality: cc-by-clean, cc-by-dirty, cc-by-sa-clean, cc-by-sa-dirty, plus a small microset and shared validation and test sets. This repository holds those as seven folders (clean, clean_sa, dirty, dirty_sa, microset, test, validation) of Parquet shards, for example 818 files in clean (measured 2026-11-14; the Hub reports about 4.8 TB). Each row holds an id, 16 kHz FLAC audio, duration in milliseconds and the transcript. Load with datasets.load_dataset('MLCommons/peoples_speech', 'clean', streaming=True) or read shards with pyarrow. This is the supervised corpus, distinct from the Unsupervised People's Speech release. MLCommons licenses the subsets under CC-BY and CC-BY-SA 4.0 for academic and commercial use; keep the subset's licence when redistributing.

## Load the data

```python
from datasets import load_dataset
ds = load_dataset("MLCommons/peoples_speech")
print(ds)
```

## Data structure

File types in sampled files: Parquet

Example file: `clean/test-00000-of-00009.parquet`

## Open it

- In a hosted notebook with the data mounted: https://scigantic.com/archive/7894b47e-8924-5292-b901-84da6b644482
- Full record as JSON (schema card included): https://api.scigantic.com/api/archives/7894b47e-8924-5292-b901-84da6b644482
- Over MCP: `get_schema_card` with id `7894b47e-8924-5292-b901-84da6b644482` 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)
