Content-based Recognition of Musical Instruments

A method for content-based audio classification is presented. In particular we focus on identification of musical instruments sounds based on timbre classification, using a biologically plausible features extraction technique called cochleagram, and a new model of recurrent neural network called LSTM. Preliminary experiments are performed to compare various feature sets and neural network sizes. In particular two experiments are performed, using two different feature sets. The best classification rate obtained is 80%, averaged on 20 trials.