Meta Unveils AI Model for Speech-to-Text Conversion Supporting 1,100 Languages

Meta announced on the 24th that it has open-sourced the ‘Massively Multilingual Speech (MMS)’ AI model, which can identify more than 4,000 spoken languages and convert approximately 1,100 languages from speech to text or from text to speech.

Meta’s MMS project aims to help many people, including those who rely on voice information, access information more easily and use various devices in their preferred languages. It is known that there are over 7,000 languages currently used on Earth, but existing speech recognition models have been limited to supporting up to about 100 languages. The AI model released by Meta this time has expanded the number of languages capable of speech-to-text and text-to-speech conversion from around 100 to 1,107, about a tenfold increase. It is significant that the number of languages identifiable by voice has increased more than 40 times, to over 4,000.

In addition to the increase in the number of supported languages, the AI model trained through the MMS project boasts superior performance compared to existing speech conversion models. According to Meta’s analysis, when the number of languages for speech conversion was increased from 61 to 1,107?about 18 times?the character error rate increased by only 0.4%. Furthermore, in a comparison with OpenAI’s speech recognition model ‘Whisper,’ the model trained with MMS data was able to handle 11 times more languages while reducing the word error rate by half.

Meta plans to contribute to the preservation of linguistic diversity and the overall development and responsible advancement of the AI ecosystem by open-sourcing the MMS model and code. It also intends to continuously expand the range of supported languages to build a single speech model supporting thousands of languages and to address dialect processing challenges that are difficult to handle with existing speech technologies. These advanced speech recognition technologies can drive innovation when applied to VR and AR technology fields or messaging services.

Meanwhile, AI is a core technology that underpins Meta’s various products and services and forms the foundation for realizing its mid- to long-term vision toward the metaverse. Meta has been focusing on three areas?building AI infrastructure, advancing AI models, and improving digital marketing experiences?while continuing research and investment in AI.