Abstract
In this thesis, three different robustness methods are used in speech recognition. The first is robustness feature extraction. CMN and DFT-MN-AUTO (Discrete Fourier Transform- Mean Normalization-relative AUTOcorrelation sequence) are tried. The second is feature parameter compensation. SBR(Signal Bias Removal), HSBRHierarchical Signal Bias Removal) and SM(Stochastic Matching) are tried. The last is model parameter compensation. BAT(Baysian Affine Transformation) and BBT(Baysian Bias Transformation) are tried.