Sequential speech segmentation based on the spectral ARMA transition measure

Abstract

Sequential segmentation algorithms based on the AR model tend to produce false alarms or to omit the change for sequences that corresponds to the ARMA model. In this paper a new sequential segmentation algorithm based on the ARMA model is presented. The ARMA model is estimated over the relatively short sequence, which has called for the implementation of the estimation algorithm with appropriately initialized starting values. The proposed algorithm adopts the MGLR concept of the sliding reference and test windows, which allow the process of decision making to be separated from the evaluation of the discrimination function. This has enabled the new triangular decision rule to be proposed; this is based on the expected shape of the discrimination function at the time of the model change. Two possible discrimination functions have been suggested. One of them is optimal in the statistical sense; the other has the better asymptotic behavior. Natural speech signal segmentation is also discussed, and an appropriate pitch-synchronous signal prearrangement has been suggested. This not only enhances the segmentation algorithm but also increases its speed, as the time can be increased by a step equal to the pitch period. The segmentation algorithm is verified on test signals as well as on the natural speech signal. The experimental results also include a comparison of the sequential AR and ARMA model-based segmentation.