theres actually a combination of things. i was doing mpeg2.5 L3 VBR files, which both had bugs, but were fixed recently

1) if i encoded it CBR but kept it mpeg2.5, it would still do the same (fixed in v3a5)
2) if i encoded it VBR and upscaled it to mpeg1, then it would lose it also (fixed in jemplode50 when it finally started to calculate the right length).

so the combination of having wrong seeks already from mpeg2.5, and wrong song length just made this thing way off. it would FF me like 7 minutes in a 80 minute lecture if i paused then unpaused. but ever since both fixes were implemented, its been fine.