A two stage mask estimation approach to robust speaker verification

We propose a two-stage mask estimation approach to robust speaker verification (SV) in noise environments. We consider a practical semi-blind SV scenario: the location of the target speaker is fixed while the locations of all interferers are unknown. In the first stage, we use a dual-microphone and a semi-blind degenerate unmixing estimation technique (DUET) to estimate an initial binary mask. In the second stage, we refine the mask based on the time and frequency histograms of the initial mask. As a result, only highly reliable time-frequency components in the spectro-temporal features are kept for downstream verification. Experiments show that the proposed approach is superior to a baseline MFCC approach and a recent local SNR based mask estimation approach.