Meta has showcased an innovative AI-powered language model, capable of recognising over 4,000 languages and transcribing speech in more than 1,100 tongues. This move is targeted at safeguarding languages on the brink of extinction.
Recognising that providing machines the skills to understand and generate speech extends information accessibility to numerous individuals, particularly those relying on voice commands, Meta has stepped up its efforts. However, crafting machine learning models for such tasks requires vast quantities of labelled data, which is limited to a few languages. The current speech recognition models merely cover approximately 100 languages, a minuscule portion of the world’s 7,000 documented languages.
Dubbed the Massively Multilingual Speech (MMS), this initiative blends wav2vec 2.0, Meta’s advancement in self-supervised learning, with a fresh dataset offering labelled data for more than 1,100 languages and unlabelled data for close to 4,000 languages. Surprisingly, the project also incorporates languages without any existing speech technology and with mere hundreds of speakers.
The models are made accessible to the public via GitHub, the popular code hosting service. Meta posits that open-sourcing these models will empower developers in crafting new speech applications in various languages, like universal messaging services or virtual-reality systems operable in any language.
Also read: Now You Can Transcribe Speech With Google Translate
Meta’s MMS models surpass existing models by covering tenfold the number of languages. The company hopes that by sharing their models and code publicly, they can contribute to the preservation of the world’s astounding language diversity.
Meta has even utilised religious texts, such as the Bible, during the model’s creation. These texts, translated into countless languages, have been extensively researched for text-based language translation studies, offering publicly accessible audio recordings of individuals reading in various languages.
Meta’s MMS project symbolises a significant stride in preserving endangered languages through speech recognition and generation technology. Meta anticipates expanding language coverage further and tackling dialect complexities, which often pose challenges for existing speech technology. Their objective is to simplify people’s access to information and device usage in their preferred language.
Meta recently launched new AI-driven ad tools on its AI Sandbox to facilitate efficient ad creation and enhance campaign results for advertisers. This platform functions as Meta’s “experimental playground” for early versions of new tools. Advertisers can generate multiple text versions, experiment with various backgrounds created from text inputs, and adjust creative assets to fit different aspect ratios across numerous platforms, saving time and resources in the process.
The news was based on a Marketing Interactive article.