ASA 125th Meeting Ottawa 1993 May

5aMU2. A parallel learning model of musical pitch perception.

Bernice Laden

School of Music, Univ. of Washington, DN-10, Seattle, WA 98195

Musical pitch perception involves the ability to extract and internalize a template that represents harmonic information. Template extraction is important to other perceptual processes such as melody recognition and chord classification, although these involve extraction of intervallic information. In order to explore how templates might be learned, a parallel learning algorithm was developed. This algorithm is based on the notion that a stimulus does not need to be physically present for a response to be learned. An artificial neural network was trained with this algorithm to identify the pitch of a tone complex. The network had 116 input units (log frequency) and 88 output units (pitch). After training, it correctly identified the pitch of 88 tone complex prototypes. It was tested with a variety of incomplete patterns as well as dyads and triads. All test patterns were correctly identified, although output strength dropped when simultaneous tone complexes were presented. The results indicate the parallel learning algorithm may be superior to other learning algorithms in terms of both accuracy and speed.