A new algorithm for voicing detection and voice pitch estimation based on the neocognitron

Abstract

Over the last decade cochlear implants have been used increasingly to restore hearing to the profoundly deaf. One of the more widely used implants is the Nucleus multi-electrode implant, developed by the University of Melbourne and Cochlear Pty. Ltd. The speech processor used with this implant is the MSP, programmed with the multipeak strategy. This device incorporates circuits to estimate the fundamental frequency (FO) of speech signals, and to decide whether voicing is present. This paper describes a new FO estimator and voicing detection algorithm based on the neocognitron; a neural network modelled on the retina and early visual system. Performance was compared with that of three other FO estimation algorithms: linear predictive coding (LPC), cepstral analysis and the algorithm used in the Multipeak-MSP processor. For the speech samples tested, the neocognitron performed more reliably than the other three systems. On the basis of these results, this work may be able to provide benefits to existing and future cochlear implant users.