AI learned to compress lossless audio 2 times more tightly

Researchers from China and Canada have introduced a new concept for data compression using artificial intelligence called LMCompress. This approach suggests that compression efficiency can be significantly increased if the algorithm understands the content of the information in the same way that humans do.

Modern users are constantly exchanging large amounts of data – text, images, videos – and need ways to speed up their transmission without losing quality. Lead author of the study, Ming Li, explains: “If a person has a good understanding of the topic, he can convey its essence in a concise form. Our approach works on a similar principle, only using large language models (LLM).”

LMCompress is based on an idea proposed back in 1948 by mathematician Claude Shannon: if the essence of information is understood, it can be transmitted more efficiently. However, it was only with the advent of AI models that can predict content that this concept gained practical implementation.

As part of the project, researchers have demonstrated that neural networks trained on large datasets are able to restore the original information with high accuracy. This opens up the possibility of not transmitting the entire dataset directly — it is enough to compress it to a level at which the model can correctly restore it to the recipient. According to tests, LMCompress doubles the compression performance compared to classical algorithms when working with texts, images and audio. As for video, the efficiency turned out to be slightly lower, but still significant.

The researchers believe that their work opens a new chapter in the development of data transmission technologies. Such algorithms can replace traditional solutions and become a standard for all digital devices.