A theoretically consistent method for minimum mean-square error estimation of mel-frequency cepstral features

We propose a method for minimum mean-square error (MMSE) estimation of mel-frequency cepstral features for noise robust automatic speech recognition (ASR). The method is based on a minimum number of well-established statistical assumptions; no assumptions are made which are inconsistent with others. The strength of the proposed method is that it allows MMSE estimation of mel-frequency cepstral coefficients (MFCC´s), cepstral mean-subtracted MFCC´s (CMS-MFCC´s), velocity, and acceleration coefficients. Furthermore, the method is easily modified to take into account other compressive non-linearities than the logarithmic which is usually used for MFCC computation. The proposed method shows estimation performance which is identical to or better than state-of-the-art methods. It further shows comparable ASR performance, where the advantage of being able to use mel-frequency speech features based on a power non-linearity rather than a logarithmic is demonstrated.