Abstract: In order to handle speech signals corrupted by noise in speaker recognition and provide robustness to systems, this paper evaluates the use of missing feature (MF) approach with a novel combination of techniques. A mask estimation based on spectral subtraction is used to determine the reliability of spectral components in a speech signal corrupted by noise. A cluster based reconstruction technique is used to remake the damaged spectrum. The recognition performance was evaluated through a speaker verification experiment with signals corrupted by white noise under different signal to noise ratios. The results were promising since they reflected a relevant increase of speaker verification performance, applying MF approach with this combination of techniques.
Index Terms: speaker recognition, missing feature.