A team of researchers has built an audio codec that steals its compression trick from large language models instead of psychoacoustics.
The system, called AudioTQ, is described in a new arXiv paper. Instead of the frequency-domain modeling that MP3, AAC, and Opus use to guess what listeners won't notice, it works directly on time-domain amplitude data. It borrows a technique from the TurboQuant framework, originally built for compressing LLM weights: a randomized Fast Walsh-Hadamard rotation that reshapes the audio signal into a predictable statistical distribution, then quantizes it with an offline-trained 6-bit Lloyd-Max quantizer plus a 1-bit residual-correction layer. The resulting 7-bit values get packed into standard 8-bit containers so the whole system runs in real time on a single CPU core, with no GPU or dedicated hardware required.
Psychoacoustic codecs work well, but they are expensive to build and tune for each domain. AudioTQ trades that complexity for math borrowed from a completely different field, and the paper reports up to 74.4 percent size reduction with a signal-to-quantization-noise ratio of about 30 dB. That is a lab benchmark, not a listening test, but it hints that compression techniques built for shrinking neural network weights might generalize well beyond LLMs.
It is still a paper, not a product, and a 30 dB SQNR is a modest quality bar next to what tuned psychoacoustic codecs already hit at similar bitrates - the real test will be whether anyone's ears can tell the difference.