Missing data imputation based on compressive sensing for robust speaker identification

In this paper, the method of missing data imputation based on the emergent field of compressive sensing for the front end of a speaker identification system in noisy conditions is investigated. Firstly, noisy speech signals are transformed into Gammatone spectrum by using cochlear filtering; then, unreliable spectral components are reconstructed given an incomplete set of reliable ones; finally, speaker features with auditory model are extracted from reconstructed Gammatone spectral data. Experimental results demonstrate that our method can improve the identification accuracy of speaker identification in noisy environments.