Modeling Time Series Signal Patterns by Statistical Distribution of Prediction Errors and its Application to Speaker Identification

Qian-Rong Gu, Tadashi Shibata

A new method that uses statistical distribution of prediction error vectors to build models of time series patterns has been developed. A universal predictor is firstly established from universal training data. Then, properties common to all the patterns are removed from the training data by the predictor. The residuals, i.e., the prediction errors, hold the characteristics of individual patterns. After clustering the prediction errors to a universal codebook, the predictor and the codebook are applied to individual training data sets to obtain the usage histograms of code vectors in the universal codebook, namely, the statistical distribution of prediction error vectors. These histograms represent the properties of individual patterns, and can be used as models in pattern recognition applications. This method is not restricted to any specific signals. As a demonstration, we utilized it to speaker identification application. It performed as well as other modeling methods under the text-dependent condition.

Full Text