SeamlessM4T is a foundational multimodal model for speech translation that enables high-quality translation between different languages. Its primary purpose is to facilitate effortless communication through both speech and text. With the increasing interconnectedness of our world and the abundance of multilingual content available, the ability to understand and communicate in any language is becoming more important than ever.SeamlessM4T supports various translation tasks, including automatic speech recognition for nearly 100 languages, speech-to-text translation for nearly 100 input and output languages, speech-to-speech translation for nearly 100 input languages and 35 output languages (including English), text-to-text translation for nearly 100 languages, and text-to-speech translation for nearly 100 input languages and 35 output languages (including English). Unlike existing systems that only cover a fraction of the world’s languages, SeamlessM4T addresses the challenges of limited language coverage and the reliance on separate subsystems by providing a unified multilingual model. It aims to bridge the gap between low and mid-resource languages and high-resource languages, improving performance for both types. Furthermore, SeamlessM4T can implicitly recognize the source languages without the need for a separate language identification model.The development of SeamlessM4T builds upon previous advancements made by Meta and others, such as the creation of the No Language Left Behind (NLLB) machine translation model supporting 200 languages and the Universal Speech Translator for Hokkien, a language without a widely used writing system.SeamlessM4T is built on the multitask UnitY model architecture, which enables the generation of translated text and speech, as well as automatic speech recognition, text-to-text, text-to-speech, speech-to-text, and speech-to-speech translations. It utilizes lightweight and highly composable tools like fairseq2, a PyTorch ecosystem library, to enhance its modeling capabilities.
