Contrary to requirements around Lossless Audio Compression Codecs, compressing audio with losses is sometimes acceptable for the benefit of accompanying lower bitrates.
There’s plenty of research and a plethora of technologies and codec implementations; can we even enumerate them, starting with the invention/popularization of MP3, or even before?
Placeholder I.
Today, we’d like to present the TSAC codec published by Fabrice Bellard just recently in 2024.
– TSAC: Very Low Bitrate Audio Compression
Please note because the reference implementation favors a GPU silicon for encoding today, it is currently not suitable for general purpose use. Please also note recent works of Fabrice Bellard are commercial software and no longer open source.
A GPU is necessary for fast operation. CPU only is also supported but slower.
tsac relies on a modified model of the Descript Audio Codec prolonged for stereo and a Transformer mannequin to additionally improve the compression ratio. Both fashions are quantized to eight bits per parameter.
The Transformer mannequin is evaluated in a deterministic and reproducible means. Hence the outcome doesn’t rely on the precise GPU or CPU mannequin nor on the variety of configured threads. This key level ensures {that a} compressed file may be decompressed utilizing a completely different {hardware} or software program configuration.