LibriVox

LibriVox coordinates volunteers all over the world who record public domain books as audiobooks and release the results for anyone to use. The catalogue runs to tens of thousands of works in many languages, from novels and poetry to non-fiction, every one drawn from a text whose copyright has expired. Because both the source books and the recordings sit in the public domain, you can use the audio freely, commercial work included, with no attribution required.

Recordings are distributed as MP3, often at several bitrates, and can be downloaded chapter by chapter or as complete works, with the matching source text usually linked from the same catalogue page. That pairing of spoken audio with a known, freely reusable transcript is the real prize for machine learning. In practice you fetch the audio and text, then run forced alignment, matching each spoken segment to its place in the written text, to produce clean audio-transcript pairs. That is exactly how derived corpora such as LibriSpeech were built.

It is at its best when you need large amounts of freely licensed speech: training or fine-tuning speech-to-text models, building text-to-speech voices, or feeding a multimodal RAG system that has to retrieve across both audio and its transcript. It doubles as a big store of narrated literature if your application reads books aloud or searches within them.

The main thing to plan for is consistency. Recordings come from many different volunteers, so audio quality, microphone setup, reading pace, and accent vary widely, and some transcripts drift from the audio. Budget for a filtering and alignment pass if you need uniform data rather than expecting a ready-made clean set.

Set against its neighbours, LibriSpeech and LJ Speech are the polished, pre-aligned datasets you might reach for first, while LibriVox is the upstream well they draw from: use it when you want more material than those curated sets offer, or need control over exactly how the audio is segmented and cleaned. Project Gutenberg is its text-only counterpart for the same public domain works.