Vocal tract length invariant features for automatic speech recognition

The effects of vocal tract length (VTL) variation are often approximated by linear frequency warping of short-time spectra. Based on this relationship, we present a method for generating vocal tract length invariant features. These new features are computed as translation invariant, correlation-type features in a log-frequency domain. In phoneme classification experiments, their discrimination capabilities turned out to be considerably better than for Mel-frequency cepstral coefficients (MFCCs). The best results are obtained when VTL-invariant (VTLI) features and MFCCs are combined. The superiority of the combined feature set and its resilience to VTL variations is also shown for word recognition, using the TIDIGITS corpus and the HTK recognizer