VoxLingua107

VoxLingua107 is a speech dataset for training spoken language identification models. The dataset consists of short speech segments automatically extracted from YouTube videos and labeled according the language of the video title and description, with some post-processing steps to filter out false positives. VoxLingua107 contains data for 107 languages. The total amount of speech in the training set is 6628 hours. The average amount of data per language is 62 hours. However, the real amount per language varies a lot. There is also a seperate development set containing 1609 speech segments from 33 languages, validated by at least two volunteers to really contain the given language.