Robust Features for Speaker-Independent Speech Recognition Based on a Certain Class of Translation-Invariant Transformations

The spectral effects of vocal tract length (VTL) differences are one reason for the lower recognition rate of today's speaker-independent automatic speech recognition (ASR) systems compared to speaker-dependent ones. By using certain types of filter banks the VTL-related effects can be described by a translation in subband-index space. In this paper, nonlinear translation-invariant transformations that originally have been proposed in the field of pattern recognition are investigated for their applicability in speaker-independent ASR tasks. It is shown that the combination of different types of such transformations leads to features that are more robust against VTL changes than the standard mel-frequency cepstral coefficients and that they almost yield the performance of vocal tract length normalization without any adaption to individual speakers.