Improved Spectral Subtraction Technique for Text-Independent Speaker Verification

The presence of different types of noise during enrollment and verification phase results in severe performance degradation in speaker verification systems. Spectral subtraction is a speech enhancement method which is often used to estimate the clean speech. However, spectral subtraction loses its accuracy in the frames with low signal-to-noise-ratio. In this paper, we present a variance measure for the low signal-to- noise-ratio frames which reflects the effectiveness of the estimate given by the spectral subtraction. This variance measure is then used to calculate the expected value of the log- likelihood score during the verification phase. Our experiments with various types of noise shows that the proposed method improves the performance of the spectral subtraction method by up to 23%.