This paper presents a new solution to the asymmetry problem of the test statistic used in speech segmentation. It is shown that the asymmetric behavior can be greatly influenced by the model identification procedures. The commonly used growing memory Burg procedure and the autocorrelation method can intensify the asymmetric behavior. Our proposed method consists of using adaptive models to accomodate to the nonstationary speech environment. These models can effectively reduce the asymmetric behavior of the test, and allow a simple on-line segmentation algorithm to be constructed. This algorithm is quite suitable for variable rate speech coding systems, and one is now being developed. This method can also be adapted to other applications, such as speech labeling and speech recognition. KEY-WORDS: speech segmentation, test statistic, asymmetric behavior, adaptive model, variable rate speech coding.
Bibliographic reference. Feng, G. / Achab, N. / Combescure, R. (1991): "On-line speech segmentation using adaptive models: application to variable rate speech coding", In EUROSPEECH-1991, 705-708.