Researchers at Meta, the company responsible for Facebook, Instagram and WhatsApp, have developed an artificial intelligence system called SEAMLESSM4T, inspired by the Babel Fish from “The Hitchhiker’s Guide to the Galaxy”.
This impressive AI is capable of translating voice or written communications into up to 101 languages almost simultaneously. It is also available in open source format for non-commercial purposes. The company explained that this AI service can perform translations in multiple ways, including speech-to-text and text-to-speech.
Meta’s AI tool trained with 4.5 million hours of recordings
Meta’s new AI, SEAMLESSM4T, recognizes speech in 101 languages, processes text in 96 languages and produces translated speech in 36 languages. Most of the training data is in English, which limits its effectiveness in languages with less data. Nevertheless, developers believe that scaling up multilingual translation systems can improve overall performance, even in less represented languages.
To train the system, 4.5 million hours of recordings in various languages were used. This amounts to about 513 years of material in total. The system was also tuned to minimize gender bias by avoiding translations that could reinforce stereotypes, such as translating gender-neutral terms into gender forms in other languages.
Traces of life have appeared on the planet Mars. NASA researchers found traces of ancient water on the red planet.
Despite these advances, translation experts warn that SEAMLESSM4T still needs careful analysis before it can be used in critical fields such as medicine and law. In these contexts, accuracy in translations is crucial and the technology can be imperfect, especially for less represented languages. This raises concerns about its application in situations requiring high reliability.
