A Novel Robust Speech Feature Based on the Mellin Transform and Speaker Normalizatuin

A novel robust feature of speech signal has been proposed by us in [1]. The new feature is the modified Mellin transform of the log-spectra of speech signal and is short for MMTLS. Due to the scale invariance property of the modified Mellin transform, the MMTLS is insensitive to the vocal tract length of different speakers. Thus it is more appropriate for speakerindependent speech recognition than the widely used MFCC. In this paper, an improved MMTLS has been proposed. The experiments show that, the improved MMTLS outperforms the original MMTLS in the performance of speech recognition. For the comparison, the frequency warping (FWP) approach based speaker normalization is also investigated. Experiments show that the performance of the improved MMTLS-based speaker-independent recognizer is much better than that of the MFCC-based one even after the latter system is combined with a technique of speaker normalization.