Skip to content

Dataset · MLCommons

50-language spoken words corpus

MLCommons/ml_spoken_words

Large-scale set of one-second audio recordings of individual words in 50 languages for keyword spotting.

What's inside

Includes 23.4 million one-second audio fragments (over 6,000 hours of recording) covering over 340,000 keywords in 50 languages, including Arabic and Czech. The data was obtained through forced alignment of crowdsourced sentence recordings, with alignment timings attached. Distributed under the permissive CC-BY 4.0 license.

What you can do with it

  • Training keyword spotting and voice command recognition models
  • Building term search systems in audio recordings for call centers
  • Developing multilingual voice interfaces for consumer devices
Tasks
Audio classification
Languages
Arabic, Assamese, Breton, Catalan, cnh, Czech, Chuvash, Welsh and 4 more
Size
10M–100M rows
Downloads this month
9.8K
Likes
37

Models trained on it

How to load it

from datasets import load_dataset

ds = load_dataset("MLCommons/ml_spoken_words")