SeamlessM4T is a voice translation product based on a multimodal model, supporting automatic speech recognition, voice translation, text translation, and voice synthesis in nearly 100 languages. This product utilizes a novel multi-task UnitY model architecture, enabling the direct generation of both translated text and speech. SeamlessM4T's self-supervised speech encoder, w2v-BERT 2.0, learns to identify structure and meaning within speech through the analysis of millions of hours of multilingual audio. The product also provides multilingual voice and text datasets like SONAR and SpeechLASER, as well as the fairseq2 sequence modeling toolkit. The release of SeamlessM4T signifies a major breakthrough in AI technology for achieving voice translation.