Abstract: | This thesis describes the development of a neural network based system for the detection and identification of instrumental pitches and timbres in recorded music. Modelling human auditory cognition, sounds are analysed using a constant Q fourier transform to give a frequency representation which is used to train and test networks of different sizes. Transient sounds are represented by multiple spectra and presented to a very simple recurrent network. Analysis of trained networks on unheard sounds demonstrates the potential of this system for inclusion in a music transcription system.
|