Also, it can help to play around with the prediction coefficients (adpcm.c - Lines 27..30, and adpcm_BlockProcess.s - Lines 154..157). Some songs have a better PSNR when there's a larger range of frequencies available (eg. Cos[(i+0.5)*Pi/16]; useful for 'noisy' stuff), and others compress better with more narrow ones (eg. Cos[(i+0.5)*Pi/32]; useful for 'smooth' stuff).