Comparison of spectral derivative parameters for robust speech recognition

Recently, spectral first-derivative parameters obtained by frequency filtering (FF) have been successfully used in both clean and noisy HMM speech recognition. In this paper, two types of spectral derivative parameters, the usual FF features and the relative spectral difference (RSD) features, are compared both between them and with their second-derivative versions. Additionally, another kind of recently introduced robust speech features, the SBCOR parameters, are related theoretically with the second-derivative RSD. By experimentally comparing all those types of features in the Aurora 2.0 noisy database framework, we conclude that the first-derivative parameters are preferable to the secondderivative ones (and to the MFCC) for both clean and noisy speech recognition, and the RSD parameters show the best average performance.