Auditory Features with Vocal Track Length Normalization for Language Identification

摘要

This paper reports on a novel feature, auditory cepstrum coefficient (ACC) with vocal tract length normalization (VTLN), for language identification (LID). The ACC feature is based on the auditory characteristics of human ear and the VTLN technology compensates the speaker variability. The detailed implementation of ACC feature with VTLN in frequency domain is given. Experimental results show that the proposed auditory feature outperforms its widely used Mel-frequency cepstrum coefficient (MFCC) counterpart and is more effective when combined with VTLN.