An eigendecomposition based two sided linear prediction model for robust speech recognition

A new feature extraction using eigendecomposition based two-sided linear prediction modelling of speech is proposed and its application to robust speech recognition is presented. The two sided linear prediction model for speech is shown to be robust against additive noise. Also the noise contamination effect can be reduced by using the reduced rank eigenvalue decomposition approach in the parameter estimation stage. In addition, a subspace noise subtraction technique is applied such that the noise level and its effect can be further suppressed. Simulation results are presented and there is a considerable improvement in the proposed new model over the conventional approaches, especially for low signal-to-noise ratio cases