A Novel Adaptive Maximum-likelihood Deconvolution Algorithm For Estimating Positive Sparse Spike Trains And Its Application To Speech Analysis

摘要

Kormylo and Mendel [101 and Kormylo [9] proposed a zero-mean Bernoulli-Gaussian (B-G) model for a sparse spike sequence /spl mu/(j) with random amplitudes such as the reflectivity sequence in seismology. Based on the zero-mean B-G model for /spl mu/(j), several adaptive deconvolution algorithms [12-14] have been reported for estimating the desired signal /spl mu/(j), which was distorted by an unknown slowly time-varying linear system v(j), with a given set of noisy measurements z(j) (= /spl mu/(j)*v(j) + n(j)). However, there are some cases that the desired sparse spike sequence has only positive amplitudes such as quasi-periodic positive pulse trains in voiced speech production model [161 and spectral pulses in spectroscopy [171. In this paper, we use a positivem-mean B-G model for positive sparse spike sequences. Then, we propose a novel adaptive maximum-likelihood deconvolution (MLD) algorithm for estimating positive sparse spike sequences from noisy measurements z(j). Some experimental results with voiced speech data are provided to show that the proposed adaptive MLD algorithm works well.