DETECTION OF PITCH IN RANDOM ACOUSTIC-SIGNALS BY NEURAL NETWORKS

By means of a backpropagation neural network a model has been built which is able to distinguish between noises and tones endowed with a detectable pitch in a (computer simulated) random acoustic environment where the information carried by the signals is compressed to its essential part by the reduction of the Fourier transform into templates of 12 numbers used as inputs. It is found that a neural network able to detect a pitch is also able to recognize the presence of a residue pitch in the signals of complex tones where the first (or the first two) harmonic has been subtracted. Finally, the correlations between the concept of consonance and the presence of a detectable pitch in the superposition of pairs of complex tones are briefly investigated.