After several months of sleepless nights and additional grey hairs, i finally got the breakthrough some days ago. It’s a clean room reverse engineered Propellerhead’s ReCycle REX2 format together with their DWOP audio encoding:
This was more of a research project to build expertise. Various LLMs helped me in the discovery process, but also result analysis, testing, fuzzing, setting up CI… Nothing would have been possible without the human in the loop 
Someone even wrote a VCV Rack module out of it: gorkulus/rex-player !
Hope it could be useful to anyone looking to add loading and saving REX in their samplers…
5 Likes
Neat !
Is there anything interresting in the dwop fromat or was it just a way to encypher the content ?
Thanks !
1 Like
I saw you post about this recently, and I am excited to add this to some apps I’ve written for doing sample management for several different eurorack samplers!
1 Like
Well, format dates back in the late 90s where most LPC or compression algorithms were not as advanced as today.
The design is a per-sample adaptive predictor bank, doesn’t have any block structure (like FLAC or WavPack), this compromises compression ratio pretty much but:
- it’s fully streaming
- it eliminates block boundaries artifacts
- the adaptive predictor amortizes its learning across the whole file
- there is no cross channel prediction beyond M/S decorrelation (both channels have independent banks and entropy so they can run in parallel).
- (but it makes random access impossible, which might be fine for the purpose it was created)
They are doing x2 integer precision tricks and zigzag sign folding. So double the samples before entering the codec and halved at the decode, this means all predictor arithmetic stays in integer without rounding, keep final halving exact.
The design clearly prioritized decode speed (simple integer ops, no division, no floating-point in the hot path) over compression ratio sensible for a real-time loop playback use case.
1 Like
