2M-Belebele

2M-Belebele Highly-Multilingual Speech and American Sign Language Comprehension Dataset We introduce **2M-Belebele** as the first highly multilingual speech and American Sign Language (ASL) comprehension dataset. Our dataset, which is an extension of the existing Belebele only-text dataset, covers 74 spoken languages at the intersection of Belebele and Fleurs, and one sign language (ASL). The speech dataset is built from aligning Belebele, Flores200 and Fleurs datasets as well as recording completely new audio for the sentences missing in Fleurs. We also provide new recordings for the Belebele question and answers as these are not in the original Flores200 dataset. Therefore, as a by-product, we also extend the Fleurs dataset (which is widely used to benchmark language identification and automatic speech recognition) by providing recordings for more Flores200 sentences than were previously available and adding sign language, creating a new **2M-Flores**. This 2M-Flores extends Fleurs by +20%. The ASL dataset is built with completely new controlled recordings of ASL signers and each flores sentence as well as questions and answers are available in video format. Speech Dataset The huggingface dataset `facebook/2M-Belebele` provides the speech version of 2M-Belebele, We will soon release the ASL version under `facebook/2M-Belebele-ASL` as it has a slightly different format (videos instead of audio). Here is a sample code to use this dataset: Columns link: the link of the original document containing the passage. question_number: the question number for this passage. Some passages have multiple questions. flores_passage: the paragraph for the passage, coming from belebele text question: the text question mc_answer1: 1st answer, text mc_answer2: 2nd answer, text mc_answer3: 3rd answer, text mc_answer4: 4th answer, text flores: details about the flores entries in this passage. A list of structs with ids and split of the original flores entry, in order of appearance in the