Meta’s new artificial intelligence can perform speech translation of 101 languages ​​of the world

Meta has recently unveiled a new artificial intelligence model that can translate 101 languages ​​at the same time and show more accuracy compared to previous models.

Classical speech translation is usually done in several steps, which in most cases starts with converting the input language into text. Next, the generated text is translated into the second language, and then the output text is converted back into spoken form in the target language. This method is known as a non-optimal path due to the number of steps, which may include errors and mistakes in translating or converting text to speech and vice versa at each step.

How is fast speech translation done by artificial intelligence?

Artificial intelligence researchers at Meta recently unveiled a new model called SeamlessM4T in an article published in the prestigious journal Nature, which, unlike classic models, can perform a more direct speech translation from the source language to the destination. The key to optimizing the new meta model lies in a process called parallel data mining. During the mentioned process, artificial intelligence looks for phrases similar to spoken input in the subtitles of videos published on the web. Next, the collected pieces are put together and after integration, its output is delivered to the user.

Meta researchers claim that the new model developed by them can show up to 23% more accuracy when translating speech compared to other models and supports 101 languages ​​of the world. It is worth noting that another similar model called AudioPaLM has been developed by Google, which supports 113 languages. Of course, Google does not provide all parts of its model in an open source form, and ordinary users face limitations in accessing it.

Despite the many advances made in the field of artificial intelligence and machine translation, researchers believe that the need for human translation is still felt strongly. Translators can understand the cultural themes inside the texts and make sure to convey their meaning in the correct way in the target language. Also, it is very important to check machine-translated texts in sensitive areas such as law and medicine to reduce the possibility of errors. Fortunately, the new Meta artificial intelligence model has been published in open source format and those interested can test its features independently.