Advances in neural network development have led to the emergence of a new class of audio encoders capable of packing audio data even more efficiently than psychoacoustic encoders. The secret to their success lies in the fact that neural codecs use not only knowledge about the characteristics of the human auditory system, but also knowledge about the structure of music and speech itself, their syntax or grammar. Neural networks acquire this additional knowledge through training on a large amount of musical and speech material. In terms of volume, this new knowledge surpasses psychoacoustic knowledge, which is why neural audio codecs require significantly greater computing resources to operate and are more cumbersome in themselves.

Their development is currently in its early stages, but many companies/developers have already presented working samples of neural audio codecs. A fairly complete list of developments can be found here - https://github.com/FORARTfe/HyMPS/blob/main/Audio/AI-based.md#codecs-

All of them exist in the form of source code and require some skill to try them out. However, there is one exception. Fabrice Bellard [https://bellard.org/] took the code from the Descript Audio Codec project as a basis, modified it for stereo signals, and compiled ready-made executable programs for Linux and Windows. This TSAC codec requires a GPU to work, but it also works on regular processors with AVX2 support, albeit slowly. The compression ratio is truly outstanding. With fairly acceptable quality, the average bitrate for 44/16/stereo material is about 7 kbit/s. For SE test samples, this bitrate turned out to be 6.37 kbit/s. This neural audio codec may not be the most perfect one at the moment, but it is available for testing by regular users and is therefore the first to be added to the listening tests on SE. An additional section, Encoders - 8 kbit/s, has been introduced for the purpose - https://soundexpert.org/encoders-8-kbps

More details about the codec can be found on its page - https://bellard.org/tsac/

As usual, anyone can participate in our listening tests. Feel free to try - https://soundexpert.org/testing-room

And as usual, it should be noted that since the testing is blind, you will not necessarily get the test file of this particular codec. After submitting your results, you will find out what you tested. Thank you in advance )!