Hierarchical AR model for time varying speech signals

The auto-regressive(AR) model is adopted to the trajectories of speech feature parameters such as linear predictors and formant frequencies. The procedure is hierarchical in its structure and is expected to be suitable for the analysis of time varying speech or non-stationary parts of speech. The method is formulated in matrix form and a feature transition matrix is introduced to express the temporal variation of feature parameter vectors. Analysis examples for CV syllables are shown and the method is confirmed by successful reconstruction of the trajectories of feature parameters based on the analysis results. The method is divided into two stages of LP analysis and the problems are to choose the preferable feature parameters for the second stage analysis and to find the appropriate values for the analysis parameters such as the window length, shift interval for the first stage analysis, the prediction order and the window length for the second stage analysis.