Reconstruction of missing features based on a low-rank assumption for robust speaker identification

摘要

Reconstruction of missing features promotes robustness in speaker recognition applications under noisy conditions. In this paper, we aim at enhancing the reliability of speech features for noise robust speaker identification under short training and testing sessions restrictions. Towards this direction, we apply a low-rank matrix recovery approach to reconstruct the unreliable spectrographic data due to noise corruption. This is performed by leveraging prior knowledge that the speech log-magnitude spectrotemporal representation is low-rank. Experiments on real speech data show that the proposed method improves the speaker identification accuracy especially for low signal-to-noise ratio (SNR) scenarios when compared with a sparse imputation approach.