Adaptive lattice analysis of speech

The adaptive lattice has been shown to have fast convergence properties. In this paper we show experimentally that the convergence of the adaptive lattice is fast enough in that its performance cannot be distinguished from that of the optimal adaptive autocorrelation method for applications in speech analysis/synthesis. A generalized window filter is used in the adaptation process. We show that while the human listener prefers window filters with an effective length of about 20 ms, the shape of the window is especially important for better speech quality. The results have implications for other applications of adaptive filtering.