Chinese Consonants Stationarity And Parameters Based On Time Series Analysis

The traditional method for Chinese speech recognition is using stationary short-time predict to stretch parameters. However, the non-stationary structure of Chinese consonants does not fit this assumption. On the other side, the large number of the parameters makes calculation difficult. How to select new parameters for recognition, especially for consonants, is one of the key directions of Chinese speech recognition research.This paper attempts to analyze and measure the stationarity of the speech signals, find new models and stretch effective parameters for recognition. First, we analyze the stationarity of 18 typical Chinese consonants with unit root tests. Secondly, sort the 18 consonants into two classes, stationary and non-stationary ones, and try different models on each class.For the stationary consonants, different ARMA models are used for recognition. The experiments show that the best two parameter groups improve the correct rate by over 12% compared to the traditional LPC(16), while the number of parameters is only 1/2 or 1/4 of the latter. For the non-stationary consonants, GARCH model is used and stationarity parameters are added for recognition. The experiments show that LPC(8) with stationarity parameters is almost as effective as traditional LPC(16), while the parameters are fewer. Moreover, the amended GARCH model is better than the traditional LPC(8), while the number of parameters is only 1/4 of the latter.The results above show that it is practicable to use stationarity for getting new recognition models and parameters. For the stationary and non-stationary consonants, it is more effective to use different models. Moreover, time-series analysis models in econometrics are worth considering for speech recognition.