Estimation of Vocal Spectra Using Maximum Entropy Method - on Number of Sample Data and Estimation Order

Abstract

In this paper, an analysis with the use of the maximum entropy method (MEM) was made for estimating voice spectra. By this analytical method, features regarding changes in the number of data and changes in the prediction error orders, in particular, were obtained from computer experiments. As a result, it was made clear that even when the number of data is small, it is possible to extract voice spectra more effective than FFT. It was further made clear that in case the prediction error order is large, there are a large number of confusing spectra so that it is not suited in such a case to extract spectra. From these results, it was found that the prediction error order of about 10 to 15 is most appropriate. It was shown that although these analyses might require somewhat long computational times, it is possible to make necessary analyses within sufficiently practical times by configuring the digital circuits by means of specialized hardware.