The speaker recognition of noisy short utterance

Abstract

The noisy short utterance is polluted by noise and its corpus is not full, so the recognition rate significantly decreased. This paper proposed noise separation algorithm based on constrained Non-negative matrix factorization (CNMF), use it to separate pure speech from noisy speech. And then the speech frames are classified to high quality class and low quality class using differences detection and discrimination algorithm (DDADA) proposed in this paper. Combining features group with GMM-UBM two-stage classification model to make full use of limited information. Experiments show that the above algorithms improve speaker recognition rate of noisy short utterance. © 2013 Springer-Verlag Berlin Heidelberg.