Many musicians require an entire band’s presence in order to effectively improve. Another fast way musicians can improve is by learning how to play or transcribe songs. Our project was aimed at making it possible to play with a band and transcribe songs faster, thereby expediting improvement. The main goal was to take a prerecorded audio track, consisting of multiple instruments playing together, and separate it into several “stem” tracks of only one instrument each. Our main attempts at this goal used non-negative matrix factorization methods, which factors the recording’s spectrogram of soundwaves to it split up into single-instrument tracks. We hoped to move on to build a mathematical model which characterizes the sound of each instrument in a “dictionary” of particular recording stems. Then, again separate an audio file into tracks by statistically matching parts of the audio file to the instrument models we have created using Bayesian statistics and Gibbs sampling. Unfortunately we were unable to fully separate a song into stem tracks, but we do have several interesting findings and suggestions for future work.