Sorry for my long absence from this thread. It's taken me a while to catch up and properly investigate the audio-stuttering problem. I am now sure that the trouble is that madplay/libmad can't keep up with decoding in real-time, as can be demonstrated by running it directly from the command line on a suitable file---the CPU is flat out and audio still sometimes breaks up. By comparison, the CPU is very lightly loaded by the original Rio application. Maybe they were using an exceptionally efficient mp3 decoder, but it seems possible there is something wrong with my build of libmad. Unfortunately, I've since messed up the toolchain I used (trying to swap from scratchbox to sb2 and back) and I'm currently having difficulty building anything at all.

In the meantime, I suggest those who can't live with limiting the mp3 bit-rate try this
kludged version 0.7b

I hope to have better results to report soon.