SFX ripping/decoding/dumping features?

[Qaazavaca Qaanic]

Trying to port over the vgmstream code to python because I can't write a wrapper around their ADPCM code.
I think that somewhere out there, someone knows the GC ADPCM algorithm by heart (really) and can reverse-engineer the coefficients just by listening to the audio (just kidding).

And here I am, struggling to even understand what you mean by 14 samples per 8 bytes... (where did the 2 missing samples go?)

vgmstream gave me the impression that it was a continuous byte stream, each 1/2 bytes corresponding to one 16-bit output sample (haven't figured it out completely).

also, is dsp (hcs, brstm) the same as dtk audio?