Taylor series expansion of psychoacoustic corruption function for noise robust speech recognition

In this paper, we present model domain adaptation algorithms for noise robust speech recognition. We have proposed a Taylor Series expansion of the psychoacoustic corruption function, which provides for superior noise robustness. The proposed joint adaptation algorithm consists of Psychoacoustic Model Compensation (Psy-Comp) and Model-domain Cepstral Mean Normalization (MCMN). While the Psy-Comp compensates for the additive noise component, the MCMN compensates for the channel component. The proposed algorithms are validated through experiments on noise corrupted TIMIT speech recognition database. We show that the joint compensation scheme provides 5% (absolute) performance gain compared to the Vector Taylor Series (VTS) scheme. Also, the proposed technique is computationally efficient and faster as compared to the VTS scheme.