Meta’s New Translation AI Is Nearly a Babel Fish

International translators in science fiction, such as Babylon in Hitchhiker guide to the galaxyHave you long offered a dream of simultaneous translation from a spoken language to another? Now, in what may be a basic step towards making this imagination a reality, scientists have developed the Meta system AI system that can immediately translate speech and text, including speech translations into direct speech, up to 101 languages.

“The science fiction provides a clear goal that our group can focus on,” he says. Marta Costa JosaMENLO Park, California Research team in MENLO Park, California. Description of scientists Their work On January 15 in the magazine nature.

As the world grows more interconnected, people have more access to multi -language content than ever. However, most automated translation systems are designed on the text of entry and output only. To date, the speech machine translation systems cover the speech that have been present in languages ​​much lower than the text systems to the text. Moreover, speech systems into previous speech were deviant towards translating a specific language into English, rather than the English language into another language.

Meta Semessm4t translation technology

Meta has now developed an AI system called SeamlessM4T that can translate speech and text in languages ​​of up to 101 languages. Specifically, you can support translation of speech into speech from 101 to 36 languages, translating speech into text from 101 to 96 languages, translating a text into words from 96 to 36 languages, translating from text to text from text for 96 languages, and identifying automatic speech of 96 languages. (Whether it can or cannot translate between languages ​​depends on the availability of quality speech data, says Costa Josa.)

For a smooth development, the researchers trained brain simulation Nerve network Artificial intelligence system is 4 million hours of multi -language sound and tens of billions of sentences from warehouses available to the public for web data. They also analyzed nearly 443,000 hours of sound with the text matching – for example, online videos with translations – to improve the system.

When it comes to translating speech to words, the research team found that Semlislangt translations were 23 percent more accurate than previous modern systems. With the tasks of speech to the text, it was 8 percent more accurate than the previous systems.

Moreover, SemlessM4T was almost 50 percent more flexible against background noise and changes in how to talk about amplifiers when it comes to speech tasks to the text. Moreover, it can translate words that mix two or more languages.

Checking toxicity and bias

To reduce the opportunities that Semlislalsm4t may add profanity and other toxic language to their translations, researchers used two strategies to remove toxicity during training and operation. When they compared Seflesm4T models with the artistic state, they found that these methods reduced toxicity in translations by up to 20 percent.

“The smooth M4T work is keen to scrutinize the possible damage in its translations,” says Alison Koenki, Assistant Professor of Information Technology at Cornell University in Ethaka, New York, who has not participated in this research. “This is especially important as the device’s speech translation is increasingly used in a set of high -risk applications, from medical appointments to employment of the workplace.”

The commanders of the commandments also studied whether Seemlessm4t is unfairly preferred one sex when translating neutral phrases between gender into sex languages. However, they found that they could not improve the performance of gender bias significantly, and they say they need to develop specific techniques to face this bias.

Along with the M4T incontinence, META released many AI systems supporting speech, text, and observations Tanel AlumäeAssociate Professor for speech therapy at Tallinn Technology University in Estonia, who did not work on smooth. Alomei and his colleagues have successfully used one of these tools “for things like identifying emotion from speech and detecting early cognitive decline – for example, Alzheimer’s – from speech,” as he noticed.

Meta is currently using SefaflessM4T to help automatically stick Videos on Instagram and Facebook. It also helps to enable Translation in the actual time of Spanish, French or Italian into English By speaking loudspeakers on Ray Ban glasses, Costa Josa says. To stimulate future research in speech translation techniques, Meta creates the code, tools, libraries and other resources associated with SEFLESSM4T smoothly available for non -commercial use.

A lot of work remains if the researchers want to make the global translator on the ground. Although it is interesting that current technology supports about 100 languages, “the number of languages ​​spoken in the world is much greater – 6500, through some estimates,” says Alomei.

This article appears on the issue of March 2025 printing as “Meta AI translation is almost a Babylon.”