Meta's Facebook AI Research Institute announced the release of XLS-R, a self-supervised learning model capable of handling diverse languages. It can currently recognize 128 languages, learning up to twice the number of languages with 10 times the amount of data compared to previous language learning models. According to the description, the XLS-R language learning model is based on the self-supervised sentence description recognition tool wav2vec 2.0 and was trained on 436,000 hours of publicly available sentences, building over 2 billion parameters to achieve a fluent translation capability across 128 languages. In the BABEL language test, XLS-R significantly reduced the error rate in translating Assamese, Tagalog, Swahili, Lao (the official language of Laos), and Georgian (used in Iran, Turkey, and other regions) and English compared to the previous version of the language model. In the BLEU language translation results, it achieved higher scores regardless of whether there was ample or limited reference material. Meta aims to recognize over 7000 languages globally through a single language learning model, reducing communication barriers between different languages and enabling users from diverse language backgrounds to communicate more fluently. Besides Meta's continued investment in language translation technology, companies like Google, Microsoft, and Amazon are also building language learning models using their data resources and leveraging artificial intelligence to achieve real-time translation, thereby breaking down language barriers between regions. For example, Google has integrated Live Translate, an AI-powered real-time translation feature into its Android operating system. This supports translation between 108 languages when connected to the cloud and can also translate specific languages offline using the phone's database. NVIDIA, on the other hand, plans to use its GPU acceleration to speed up the training and learning of multiple languages.
