Speaker recognition based on transformed line spectral frequencies

Abstract

Line spectral frequencies (LSF) and five types of transformed LSF are studied for robust text-independent speaker identification. Transformations are constructed by considering physical aspects of the vocal tract. These aspects are: location of formants/nulls; bandwidth of formants/nulls; bandwidth and location of formants; bandwidth and location of nulls; interval of adjacent formant and null locations. Identification tests using the TIMIT database verify that all features are useful for speaker recognition; the bandwidth and location of formants, especially, show the best performance. Simulation results also show that LSF and some of the transformed LSF give better performance than Mel-frequency cepstral coefficient (MFCC).