Time-varying analysis using ARMA model with double inputs

The authors introduce an adaptive speech analysis using a time-varying ARMA (autoregressive moving average) model with double inputs (an ARMA-D model). The method presented is an extended version of a model identification system (MIS). For improving the ability of the time-varying parameter estimates in MIS, the ARMA-D model is employed to obtain an accurate speech production model. In addition, some assumptions are introduced to represent the rapid changes of speech spectra, and a weighted window in adaptation is used for the approximation of low time-varying spectra. Experimental results show that the extended method can analyze more accurate spectra of real speech than conventional methods