VoxPopuli
VoxPopuli is a large-scale multilingual speech corpus released by AI at Meta in 2021, built from recordings of European Parliament plenary sessions and events held between 2009 and 2020. Its headline component is around 100,000 hours of unlabelled speech across 23 languages, intended for self-supervised and semi-supervised pretraining of speech representation models. Alongside this sit roughly 1,800 hours of transcribed speech in 16 languages for automatic speech recognition, and a set of speech-to-speech interpretation pairs that take the aligned total to about 5,100 hours. A separate accented-English subset provides transcribed non-native speech for robustness research.
You access it through the Hugging Face Hub, loading a single language by configuration name with the datasets library, selecting a combined multilingual configuration, or picking the accented-English split; streaming is supported for the larger partitions. Audio is decoded at 16 kHz and the Hub stores the material as Parquet, with sentence-level transcripts provided for the labelled portions.
For RAG and AI work, VoxPopuli is training and evaluation material for the speech-to-text stage rather than a text knowledge base in itself. It is best for pretraining or fine-tuning multilingual speech models, benchmarking European-language ASR, and producing transcripts that you later chunk, embed and retrieve. Because the audio is parliamentary speech, it skews towards formal, prepared delivery rather than spontaneous conversation.
The licence position needs care. The Hugging Face release is tagged CC0-1.0, which would waive rights over Meta's packaging, but the card also carries an 'other' tag and points to the European Parliament's legal notice covering the raw audio. Treat the underlying recordings as subject to that notice and check it before any commercial redistribution, which is why we flag the content licence as differing from the headline. If you already use Common Voice or Multilingual LibriSpeech, VoxPopuli complements them with far larger unlabelled coverage and a distinctly formal, European-institutional register.
Related sources
Emilia
A large-scale, multilingual in-the-wild speech dataset for speech generation, starting at over 101,000 hours across six languages and extended to more than 216,000 hours in Emilia-Large. Built for training natural, spontaneous TTS and speech-understanding models.
Granary
An open multilingual speech corpus from NVIDIA covering roughly one million hours of audio across 25 European languages, built for automatic speech recognition and speech translation.
LibriSpeech
A collection of 1,000 hours of read English speech, released under CC BY 4.0 and stored using the open-source FLAC audio encoder. Labels are aligned at the sentence level. Derived from LibriVox public domain audiobooks, it is the standard benchmark for English automatic speech recognition.
LJ Speech
An open dataset of 13,100 short audio clips of a single speaker reading passages from seven non-fiction books. Every clip is transcribed and clips run from 1 to 10 seconds. Released into the public domain, it is the standard single-speaker text-to-speech benchmark.