Channel robust speaker verification via extended feature mapping

An extended feature mapping (EFM) approach for channel robust speaker verification is proposed in this paper. This approach seeks to map the distorted feature vectors into a channel-independent space so that each individual speaker model can be trained with channel-independent feature vectors. EFM consummates the existing feature mapping idea by defining a more rational mapping function. Experimental results on closed-set speaker verification tasks on a channel-corrupted subset of TIMIT database are presented in terms of DET plot (see fig. 4). The new approach is shown to provide significant performance over baseline system, and is more stable and better than the original feature mapping and Hnorm.