French mathematician Fabrice Bellard, who founded projects such as QEMU, FFmpeg, BPG, QuickJS, TinyGL, and TinyCC, has published a sound encoding format called TSAC along with associated tools for compressing and decompressing audio files. The format is designed for data transmission at very low bitrates, for example, 5.5 kb/s for mono and 7.5 kb/s for stereo, while maintaining acceptable quality for music and speech. Using TSAC allows a musical composition lasting 3.5 minutes with a sampling rate of 44.1 kHz (stereo) to be packed into a file size of 192 KB, which will be almost indistinguishable from the original to the ear of an untrained listener. The project code is distributed under the MIT license.
The TSAC codec is based on the Descript audio codec, which has been enhanced to support stereo sound and adapted to use a different machine learning model based on a transformer architecture. This has allowed for an increased compression rate by reconstructing lost details based on the human auditory perception model. The model occupies about 200 MB in compressed form and is presented deterministically, ensuring the same result regardless of the CPU/GPU and the number of threads used in computations.
The encoder can operate using only the CPU for calculations (with AVX2 instructions supported for acceleration), but for achieving high performance, using a GPU is recommended. Currently, CUDA API can be used for acceleration with NVIDIA GPUs based on Ampere, ADA, and Hopper microarchitectures (RTX 3090, RTX 4090, RTX A6000, A100, and H100) that have at least 4 GB of video memory. FFmpeg is used for converting audio files before encoding.
Additionally, it's worth noting the update of the tool ts_zip developed by Bellard, designed for efficient text data compression, using a token prediction mechanism based on the RWKV 169M v4 large language model and machine learning system. When compressing the Wikipedia archive, the ts_zip tool achieved a compression ratio of 7.3, while compressing the Linux kernel code resulted in a 7.8 compression ratio. In comparison, the compression ratios using the xz utility were 4.7 and 5.5, respectively. The trade-off for such high compression efficiency is a low compression speed and significant resource requirements (a minimum of 4 GB of RAM). On a system with an RTX 4090 GPU, the compression performance is approximately 1 MB/s.
Source: opennet.ru
