Harmonic Structure Features for Robust Speaker Recognition against Channel Effect

This paper proposes a novel feature set for robust speaker recognition, which is based on the harmonic structure of speech signals. Channel modulation effects are supposed to be weakened in the harmonic structure features, and furthermore the influence introduced by channel variability could be diminished to a certain degree. Though experiment results show that the raw performance of the harmonic structure features is not better than our baseline system due to some inevitable information lost, significant improvement of average 9.85% relative minCost decrease has been made by a score fusion approach with baseline features.