We present a novel method for speaker identification in noise mismatch conditions. The proposed method is based on Jump Function Kolmogorov (JFK), - a new stochastic instrument ., which is (a) additive so sum of signal and noise yields the sum of their JFKs; (b) sparse so signal and noise have better separable supports in JFK's representations. The separability of signal's and noise's representations is the main advantage of JFK to make this instrument more robust in the classification then the conventional probability density function (PDF). In the approach, we develop a speaker identification system based on JFK analysis in the wavelet domain, i.e. the JFKs are estimated in each subband to match the nearest from trained templates. The experimental results show that the proposed method is comparable to the conventional method under clean condition but significantly outperformed them under noise mismatch conditions.
Bibliographic reference. Tran, Huy Dat / Li, Haizhou (2008): "Speaker identification in noise mismatch conditions based on jump function Kolmogorov analysis in wavelet domain", In INTERSPEECH-2008, 1469-1472.