Feature Enhancement: Fabrice Bellard's transformer(s)

[clort81]

From the page:
https://bellard.org/tsac/

tsac is based on a modified version of the Descript Audio Codec extended for stereo and a Transformer model to further increase the compression ratio. Both models are quantized to 8 bits per parameter.
The Transformer model is evaluated in a deterministic and reproducible way. Hence the result does not depend on the exact GPU or CPU model nor on the number of configured threads. This key point ensures that a compressed file can be decompressed using a different hardware or software configuration.

Could the authors comment on this work and evaluate it?