Feature extraction based on DCT and MVDR spectral estimation for robust speech recognition

This paper proposes a new noise robust feature extraction method for speech recognition. It is based on the discrete cosine transform and minimum variance distortionless response (MVDR) methods of spectrum estimation and differential power spectrum technique. The large bias drawback of the periodogram method can be solved by using DCT instead of FFT. The MVDR method can also increase the robustness of the features by reducing the variance of the estimated power spectrum. The above method, when evaluated on Test set A of Aurora 2 task, gave a relative improvement of up to 63.3% in recognition accuracy in comparison with MFCC as the baseline.