A software method to simulate the experience of playing a musical instrument in a small informal group (a session) would be useful for aspiring musicians who wish to learn by ear rather than by reading formal music. Such music is usually played with some variation every time, variation that is problematic for simple pattern matching algorithms. The liquid State Machine (LSM) is a recent approach to solving time-based pattern matching problems, by transforming them into a spatial representation and then applying neural network techniques. LSMs are notably resistant to temporal noise in the patterns that they are trained to identify. This thesis includes an investigation of tuning possibilities for LSM architectures, including the study of performance changes when varying single and multiple LSM parameters. The results of these studies are then applied to the problem of recognising musical phrases in real time, with and without errors in melody.