Hard-Mask Missing Feature Theory for Robust Speaker Recognition

Compared with conventional full-band speaker recognition systems, Advanced Missing Feature Theory (AMFT) produces a much lower error rate, but requires increased computational complexity. We propose a weighting function for the score calculation algorithm in AMFT. The weighting function is estimated by calculating the number of reliable spectral components. A modified mask is also proposed to reduce the number of reliable components based on the estimated weighting function. In the proposed Hard-mask MFT-8 (HMFT-8), only 8 elements are selected out of 10 spectral components in a feature vector. Compared with the full-band system and the AMFT, the proposed HMFT-8 gives a lower identification error rate by 16.95% and 2.67%, respectively. In terms of computational complexity, AMFT and HMFT-8 require 307 and 41 arithmetic and conditional operations for each frame, respectively.