Meta demonstrates speech-to-speech translation software

Meta has put a demo online of an AI model that can convert spoken text in multiple languages ​​into text and speech in another language. The model can handle it if a speaker switches languages ​​in the middle of a sentence.

The SeamlessM4T model is available as online demo, download on GitHub and demo on Hugging Face, reports Meta. It can use speech and text as input and then transcribe and translate speech or text into a combination of both. It supports speech input in nearly 100 languages ​​and output to text in 35 languages.

It comes from the SeamlessAlign dataset containing 270,000 hours of speech and text fragments with translations. Meta’s researchers consider the model an advancement because of its support for many languages ​​and the ability to have speech and text as input and output. The model is now available to researchers and developers. It is unclear whether Meta will incorporate the translation model into its own apps and services, such as WhatsApp, Instagram and Facebook, at a later stage.