An investigation into better frequency warping for time-varying speaker recognition

Performance degradation has been observed in presence of time intervals in practical speaker recognition systems. Researchers usually resort to enrollment data augmentation, speaker model adaptation, and variable verification threshold to alleviate the time-varying impact. However, in this paper, efforts have been made in the feature domain and an investigation into better frequency warping for the target task has been done. Two methods to determine the discrimination sensitivity of frequency bands are explored: an energy-based F-ratio measure and a performance-driven one. Frequency warping is performed according to the discrimination sensitivity curves of the whole frequency range. Experimental results show that the proposed features outperform both MFCCs and LFCCs, and to some extent, alleviate the time-varying impact on speaker recognition.