AI is becoming more worldly.

On Monday, Meta released an open source “omnilingual” speech recognition model, capable of supporting more than 1,600 languages, including 500 “low-resource languages never before transcribed by AI.” The polyglot model extends far beyond OpenAI’s Whisper model, which supports 99 languages.

In addition to the large swathe of languages this model already knows, Meta’s Omnilingual system can extend to new languages without requiring expert fine-tuning, relying instead on a handful of in-context examples, the company noted.

The goal is to make “spoken language universally accessible,” Meta said in its announcement. Speech recognition models, however, often focus on a "limited set of high-resource languages that are well represented on the internet.”

“This means high-quality transcriptions are often unavailable for speakers of less widely represented or low-resource languages, furthering the digital divide,” the company noted.

And Meta isn’t the only one seeking to use AI to bridge conversational gaps. In December, a government agency in the Philippines released a project called ITanong, a chatbot capable of English, Filipino and Taglish. And in October, a UK nonprofit developed an AI chatbot called Ulangizi AI specifically designed for farmers in Malawi to get advice in their own language.

But the capabilities of these models could put translators’ jobs at severe risk: A survey released by Microsoft in July that measured the exposure of occupations to generative AI named interpreters and translators as the job with the highest amount of “AI applicability.”

However, while these models have a strong grasp on language itself, communication goes beyond words. For example, AI models may struggle to convey emotion, translate words that don’t have exact equivalents across languages, understand sarcasm or recognize local slang that isn’t universal. Replacing human interpreters with AI runs the risk of meaning being lost in translation.