Granary

Granary is a large-scale open speech dataset released by NVIDIA, in collaboration with researchers at Carnegie Mellon University and Fondazione Bruno Kessler. It gathers around one million hours of audio across 25 European languages, split roughly between speech recognition (about 650,000 hours) and speech translation (about 350,000 hours). The label data is largely pseudo-labelled: audio drawn from public sources is transcribed and translated with an automated pipeline, then filtered for quality. Coverage extends to lower-resource languages such as Croatian, Estonian, Latvian, Lithuanian, Maltese and Slovenian, which are usually thin in open speech collections.

Access is through the Hugging Face dataset card, where the data is published in JSON and Parquet, organised by language and task. You can stream it or pull it in bulk with the Hugging Face datasets library. It is the training data behind NVIDIA's Canary and Parakeet speech models, so it pairs naturally with those if you want a reference point for expected quality.

For AI work it is best suited to training or fine-tuning multilingual ASR and speech translation systems, and to building spoken-language retrieval or transcription pipelines that feed a RAG system once the audio is turned into text. It is not a ready-made text knowledge base: you are working with transcripts and translations tied to audio, so expect a preparation step before indexing.

Watch-outs: the labels are machine-generated rather than human-verified, so some transcription and translation noise is inherent, and the dataset viewer has shown schema mismatches on certain language subsets. The compilation and its annotations are offered under CC BY 4.0, but the underlying audio comes from public sources that carry their own terms, so confirm provenance before any sensitive use. If you need alternatives already in this directory, consider Mozilla Common Voice for crowd-sourced multilingual speech, or the People's Speech corpus for large-scale English ASR.