GARCH Coefficients as Feature for Speech Recognition in Persian Isolated Digit
Mohammad Javad Abdolahi, Hamidreza R. Amindavar · 2006
This paper describes a new technique that gives high performance based on GARCH (generalized autoregressive conditional heteroskedastic) time series modeling incorporating past variances to predict future variances. This is particularly suitable since no transformation on the speech signal is performed, rather we have a new statistical feature extraction, moreover, the speech signals are among nonstationary processes whose variances are heteroskedastic; e.g., time varying. Therefore, we provide a new parametric speech modeling using GARCH coefficients. The features resulting from GARCH modeling are used for recognition of isolated digits 1 to 10 in the Persian language. The results show a significant improvement in the recognition accuracy compared to results based on Mel-frequency cepstrum coefficients (MFCC).