Skip to content

Datasets

Hugging Face datasets with plain-language descriptions: what's inside and what models to train on them. 4,788 in the catalog so far, with new ones added every hour.

Data type

For which models

Find a dataset for your goal

Answered by Gemini 4 · the chat is kept only in this tab

Datasetespnet/yodas3Top pick

YODAS v3 100+ language speech corpus

Over 1.1 million hours of web audio in 100+ languages for ASR, TTS, and audio representation learning.

  • Audio → audio
  • Automatic speech recognition
  • Text → speech
253.4K 2631M–10M rows
Datasetespnet/yodas3.1Top pick

Audio corpus in 100+ languages

A massive dataset of over 1.1 million hours of audio across 100+ languages for speech tasks.

  • Audio → audio
  • Automatic speech recognition
  • Text → speech
239.6K 2761M–10M rows
Datasettwangodev/librivox-mirrorTop pick

Multilingual LibriVox Audiobooks Mirror

Over 133,000 hours of voiced books across 86 languages for training speech recognition and synthesis systems.

  • Automatic speech recognition
  • Text → speech
  • Multilingual
23K 0100K–1M rows
DatasetFBK-MT/moselTop pick

EU multilingual speech corpus

Up to 950K hours of open speech across 24 official EU languages with automatic transcripts.

  • Automatic speech recognition
  • Text → speech
  • English
  • Bulgarian
  • Croatian
  • +9
8K 941M–10M rows
Datasetnvidia/GranaryTop pick

Granary: speech and translation in 25 languages

A large open speech dataset across 25 European languages for ASR and speech translation.

  • Automatic speech recognition
  • Translation
  • Bulgarian
  • Czech
  • Danish
  • +9
4.8K 238100M–1B rows
DatasetMohammadJRanjbar/ParsVoiceTop pick

Multi-speaker Persian speech corpus for TTS and ASR

Largest open audio-text corpus in Persian, collected from audiobooks for speech synthesis and recognition.

  • Text → speech
  • Automatic speech recognition
  • Audio classification
  • Persian
2.6K 451M–10M rows🔒 on request
Datasetnvidia/Granary-HiFi

NVIDIA clean multilingual speech metadata

Filtered metadata for a subset of VoxPopuli (10 to 100 million entries) for selecting high-quality audio in ASR.

  • Automatic speech recognition
  • Bulgarian
  • Czech
  • Danish
  • +9
2K 1110M–100M rows