A Predictive Minimum Description Length (MDL) Approach to Speech Clustering and Segmentation (NOT PUBLISHED)

The method has been applied to the problem of statistical clustering of speech central coefficients into Gaussian mixtures. When the data has a temporal order, e.g. speech frames in continues speech, the prediction error, as a function of time, can be used for segmentation, as well as an alternative cost function for statistical modeling. The average prediction error approaches a limiting function of time, asymptotically independent of the particular modeling scheme, and thus is a characteristic of the speech signal itself.