The effects of filter frequency scale variability on speaker identification performance

Abstract

Extracting discriminatory feature vectors that contain speaker specific information is of crucial importance in speaker identification. Although the cepstrum coefficients on the Mel frequency scale are commonly used as feature vectors, it is demonstrated in this paper that linear and ERB frequency scales provide better results compared to the Mel scale. In the paper, ERB, Bark and linear scales are compared with Mel scale on the TIMIT and NTIMIT databases. On the TIMIT database, an identification rate of 100% is obtained with the linear frequency scale when the filter-bank is placed in 0-8 KHz range, and a rate of 98.81% is obtained with the ERB scale using 0-4 KHz filter-bank frequency range. On the NIMIT database, 73.51% identification rate is achieved with linear scale, resulting in 2.97% improvement over that of the Mel scale.