Meta AI Model: Capable of Transcribing and Translating Nearly 100 Languages
Meta AI Model: Capable of Transcribing and Translating Nearly 100 Languages
Meta AI Model: Capable of Transcribing and Translating Nearly 100 Languages
On August 22, Meta released an artificial intelligence (AI) model called SeamlessM4T, which can translate and transcribe nearly 100 languages.
According to Meta, SeamlessM4T can translate speech to text and text to text in nearly 100 languages.
For speech-to-speech and text-to-speech operations, it can recognize 100 input languages and convert them into 35 output languages.
SeamlessM4T has been released under the Creative Commons License 4.0, allowing researchers to iterate upon it.
In addition to SeamlessM4T, Meta also released metadata for its open translation dataset called SeamlessAlign.
Meta stated, “Building a universal language translator, like the fictional Babel fish in ‘The Hitchhiker’s Guide to the Galaxy,’ is challenging because existing speech-to-speech and speech-to-text systems cover only a small fraction of the world’s languages.”
“The Hitchhiker’s Guide to the Galaxy” is a series of science fiction novels written by British author Douglas Adams, and the Babel fish is a fantastical creature created in this work, small enough to fit in a person’s ear and capable of living off brainwaves.
When placed in the ear, it allows individuals to understand any language.
For the SeamlessM4T model, Meta researchers stated in a research paper that they collected audio training data from 4 million hours of original audio, sourced from a publicly available web crawl data repository, although the specific repository was not disclosed.
The research report mentioned that text data was extracted from datasets created last year, which gathered content from Wikipedia and related websites.
Meta emphasized that SeamlessM4T represents a significant breakthrough as this model can complete the entire translation task in one go, unlike other large translation models that segment translation into different systems.
SeamlessM4T builds upon Meta’s previous translation models. Last year, Meta released a text-to-text translation model supporting 200 languages.
They developed datasets for multilingual speech-to-speech translation and a large-scale multilingual speech recognition dataset.
Meta showcased its universal speech translator last year, which can translate Min Nan Chinese into English.
