Includes kingshriek's patch above plus I rewrote ADPCM (including Deunan's hint that the samples go 10 32 54 76 instead of 01 23 45 67). Things still get crunchy due to DC offset drift (it's quite noticeable in AO with the 'scope) so it may be our actual decode constants aren't quite right, but in general it sounds far better.
For end users following along at home, you can hear the current code playing Sakura Taisen 4 here:
