Performance Comparison of Two Joint Three States Segmentation and AR Modelling Algorithms

摘要

Two procedures designed for the joint segmentation and AR modelling of quasi-stationary signal-the Rate/Distortion Algorithm (RDA) and the Maximum Likelihood Algorithm (MLA) - are compared regarding their performance. Both algorithms share a same parameters estimation process and differ only in the form of the criterion to be optimised. The comparison is achieved with 2 simple indexes on the phonemes Otago data corpus assuming a 3 state model of the same order for each phoneme. As a general result, it is shown that the MLA is less sensitive to AR order choice and perform the least model residual correlations.