Multilingual LibriSpeech (MLS)

Multilingual LibriSpeech (MLS) is a large read-speech corpus built by researchers at Meta AI, extending the English-only LibriSpeech into eight languages. Like its sibling, it is derived from LibriVox, the volunteer-read public domain audiobook project, with the audio segmented, aligned to its text, and filtered so that each clip tracks its transcript closely. It covers English, German, Dutch, Spanish, French, Italian, Portuguese and Polish, and totals roughly 50,000 hours of audio, though the balance is heavily skewed: about 44,500 hours are English, leaving around 6,000 hours spread across the other seven languages, with Polish and Portuguese by far the smallest.

You can get the data two ways. The original release lives on OpenSLR as per-language archives you download in bulk, with the audio stored as FLAC at a 16 kHz sampling rate. Meta also publishes a streamable version on the Hugging Face Hub, loadable per language through the datasets library with optional streaming, where the audio is served as Parquet-packed Opus. Both carry sentence-level transcripts alongside the audio.

For RAG and AI work, MLS is less a knowledge base than training and evaluation material for the speech-to-text stage that turns audio into the text you later chunk, embed and retrieve. It is best for building or benchmarking multilingual automatic speech recognition, fine-tuning models on non-English read speech, or producing transcripts for a voice-driven pipeline in several languages at once. As with LibriSpeech, the audio is clearly enunciated audiobook reading, so strong results here do not promise strong performance on spontaneous, accented or noisy speech.

The licence is clean: the corpus is released under CC BY 4.0, so commercial use is permitted provided you attribute, and the underlying LibriVox recordings are public domain. There is no share-alike obligation. If you already list LibriSpeech, treat MLS as its multilingual counterpart, and reach for the English-only set when you need the standard single-language ASR benchmark.