Using long-term information to improve robustness in speaker identification

Abstract

In this paper we propose two new methods of improving the robustness of Automatic Speaker Identification systems. These methods rely on using long-term information in the speech signal to improve the robustness of the features. The first method involves averaging filterbank parameters from consecutive short-time frames over a longer window. The second method investigates the use of frame lengths longer than generally assumed stationary. We show that these two methods result in an improvement over standard Mel Frequency Cepstral Coefficients in the presence of additive white Gaussian noise in speaker identification applications. Furthermore, additional improvements are observed at mid-range SNR when the proposed methods are used in combination.