The objectives of this work is to present a new feature representation for speech recognition in noisy environment using a combination of the Two-Sided Linear Prediction and Singular Value Decomposition (TSLP-SVD). The two-sided prediction model has very robust feature against additive noise. Furthermore, the noise effect on estimating the feature vector can be reduced by using SVD to extract the dominant components of the speech signal. A new Nonlinear Noise Subtraction (NNS) is used for the further reduction of noise level during parameter estimation.
Keywords: Noisy speech recognition, Two-Sided Linear Prediction, Singular Value Decomposition