Meta, on 23rd August 2023, released an Artificial Intelligence (AI) Model that has the ability to transcribe speech in different languages. This new development is seen as a potential building block for tools facilitating real-time communication across different languages. In a blog post, Meta noted that its SeamlessM4T model had the ability to support translations between speech and text for more than 100 languages. It also supports speech-to-speech translations for over 35 languages. This combines technology that was previously used only in separate models.
Meta CEO Mark Zuckerberg said that he sees such tools as having the ability to facilitate interactions between users from around the world in the metaverse. He added that he envisions the company having a set of interconnected virtual worlds.
This new AI model will be available to the public for non-commercial use. This year, Meta has released several free AI models, including the Llama, which is a large language model. The Llama has posed a very serious threat to similar models sold by MSFT.O OpenAI and Alphabet’s. (GOOGL.O) Google.
Mark Zuckerberg noted that the open AI ecosystem would work to the advantage of Meta since the company would gain more from crown-sourcing the creation of consumer-facing tools for its social media platforms. This is the reason why they opted for the free models. However, the creation of these models has not been smooth because Meta is facing similar legal questions as other creators of these models on the training data used to create the models.
Meta and OpenAI are already facing a copyright infringement suit filed by comedian Sarah Silverman. Sarah accused both companies of using her books as training data without permission.
In this new SeamlessM4T model, Meta researchers noted that they had audio training data. They added that the audio training data was 4 million hours of raw audio originating from a publicly available respiratory of crawled web data. They, however, didn’t specify which respiratory they had obtained this data, nor did they indicate whether they would make it available in the future. When Meta spokesperson was asked about the provenance of the audio data, he opted not to give a response on the same.