Artificial intelligence works especially well in language-related tasks. Meta – owner of Facebook, Instagram and WhatsApp – knows this well: the company released on Tuesday (22) a model called Seamless M4T, which is capable of translating and transcribing almost 100 languages.
Seamless M4T works with text and language, with several possibilities:
speech recognition for almost 100 languages;
speech-to-text translation for nearly 100 input and output languages;
speech-to-speech translation for nearly 100 input and 36 output languages;
text-to-text translation into nearly 100 languages;
text-to-speech translation for almost 100 input and 35 output languages.
To create the model, Meta says it collected tens of billions of written sentences and 4 million hours of publicly available speech on the web.
Paco Guzmán, a researcher at Meta, told Axios that Seamless M4T does not use intermediate models and even accepts mid-sentence language changes.
“This allows it to work better with populations that naturally mix languages on a day-to-day basis,” Guzmán said.
The new Meta template is not available for anyone to use. It has been released under a Creative Commons license, for non-commercial use. Researchers and developers will be able to use the technology in their work.
Meta and competitors invest in AI for translation
This isn’t the first time Meta has developed artificial intelligence models to work with translation.
In 2022, the company launched No Language Left Behind, for text-to-text translations. He was able to translate 200 languages.
Then came the Universal Speech Translator, which worked with speech in and out.
Seamless M4T builds on Massively Multilingual Speech, Meta’s framework for speech recognition, language identification and speech synthesis for more than 1,100 languages.
Among the competitors, Google is working on the Universal Speech Model, an initiative to understand the world’s 1,000 most widely spoken languages.
Mozilla is leading the Common Voice project, a collection of voices from several languages, to train speech recognition algorithms.