I started to optimize the subband synthesis with some mild success. If i use psyco and a mono mp3 at 128kps, it 10x slower than realtime. I think I’m going to have to resort to numpy tricks. A short term goal is 100% python realtime decode of an mp3 (super common python modules are fine).
Portalfire's Blog 
Pingback: Decoding MP3 in Python - Hack a Day
You might want to try Cython (or, even better, ShedSkin).
i’ve found the overhead of using numpy to not be worth it unless you’re doing TONS of matrix stuff. but good luck.
You said ” but from a learning/documenting perspective, it is quick and easy.” So, obviously you want this project to aid for learning. Then I suggest not to go for optimizations. Or instead you can have different code packages for different stages. I mean a code package for un-optimized mp3 decoder , a code package for optimized. That way it helps the learner more.