The KanNon System

Applying Minimum Cross Entropy Method Combined with Burg's Method

Abstract

We develop the real-time speech visualization system called ``KanNon" which supports speech communication of hearing-impaired people. The KanNon system presents information of the speech such as characters including voice volume, pitch by speech recognition and sound spectrogram in real-time. We apply novel spectral estimation algorithm combinig Burg's method and minimum cross entropy method (Burg-MCEM) to the KanNon system. The Burg-MCEM estimates model parameters using prior information of AR model parameter based on the MCE principle. Further, we propose Burg-MCEM with distinction of prior information using distance of the Kullback information.