by Google
FLEURS is Google's multilingual speech dataset covering 102 languages, built as the speech version of the FLoRes machine translation benchmark. It supports speech recognition, language identification and speech-text retrieval.
DatasetFree
Step 1: Install the Hugging Face datasets library.
Step 2: Load one language with load_dataset("google/fleurs", "hi_in", split="train"), using the config name for your language.
Step 3: Use the "all" config to download every language for multilingual fine-tuning or language identification.
Step 4: Add streaming=True to read samples without downloading the whole dataset.
Step 5: Use the audio and transcription fields to fine-tune or evaluate your model.
102 languages, grouped into seven geographical areas.
The card lists about 350 GB of disk used in total. A single language config is much smaller, so loading one language at a time or streaming is practical.
The card describes speech recognition, language identification and speech-text retrieval.
Each config has a train split of around 1000 examples and validation and test splits of around 400 examples each.