The People's Speech

The People's Speech comes from MLCommons, the industry consortium behind the MLPerf benchmarks, and it was built to fix a specific problem: most large speech datasets carry terms that make commercial use difficult. It collects more than 30,000 hours of English audio from openly licensed sources, largely recorded speech gathered from the web, and pairs every clip with a transcript. Audio arrives as FLAC (a lossless compressed audio format), with transcripts and alignment information in JSON.

The dataset is hosted on the Hugging Face Hub, so the datasets library is the natural entry point. Because the full corpus is large, stream it rather than downloading everything up front: streaming hands you batches of audio-and- transcript pairs on demand, which is plenty for prototyping before you commit disk space to it.

Its home turf is training automatic speech recognition (ASR), the task of turning spoken audio into text. For retrieval work specifically, the transcripts are the asset. If you are building a system that searches over spoken material such as podcasts, lectures, or recorded calls, People's Speech gives you realistic transcribed speech to test chunking and retrieval against, and the paired audio lets you evaluate a full speech-to-retrieval pipeline from sound to answer.

Two things deserve caution. Because the transcripts were assembled automatically at scale from varied sources, their quality is uneven: expect misalignments and errors, which makes the set better for training and stress- testing than as a pristine reference corpus. And the licence is split. Some subsets are CC BY 4.0 and others CC BY-SA 4.0. Both permit commercial use and require attribution, but the share-alike (SA) portions require any derivative database you publish to carry the same licence, so track which subset each clip came from.

For clean single-speaker audio aimed at text-to-speech, LJ Speech is the usual pick, and for broad multilingual coverage OpenSLR and its LibriSpeech collections are worth a look. Reach for People's Speech when permissive commercial licensing and sheer scale matter most.