Nonlinear translation-invariant transformations for speaker-independent speech recognition

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