Dynamic Robust Speech Recognition

Robust recognition theory has become one of research focuses of acoustic speech recognition. Acoustic speech digital signal is a random process repeatedly alternating stationary pieces with non-stationary pieces. However both the current linear and stationary characteristic parameters drawn from such signals and the rigid recognition models do not adapt to such repeatedly alternating property of acoustic speech. Though Missing Feature Approach (MFA) has been proved a considerable solution of enhancement of robustness for noisy speech, MFA classifying in binary way seems to be rough and it cannot used to deal with cepstral feature. Consequently, current noisy speech recognition systems perform mostly poorly. This paper tries to set up dynamic recognition theory that applies non-linear doubly random time series instead of linear model and auto-select types of parameters and recognition models based on nonstationary measure of real-time. Meanwhile this theory gives two approaches of Feature with Confident Weight (FCW) in three means to describes the effect of noise in a more precise way and available in cepstral domain. Experimental results show proposed approaches could improve the recognition accuracy significantly in adverse environment, including stationary and non-stationary noise environments.