An affine transform for speaker recognition enhancement under mismatched coding conditions

Text-independent speaker recognition performance suffers significantly under mismatched coding conditions between training and testing speech data. In this paper, a baseline HMM-based speaker recognition system is tested under various mismatched conditions with a large number of different HMM topologies. Training and testing the models using only the voiced segments of the samples is then considered. A technique based on a diagonal affine transform in the cepstrum domain is proposed, which maps the mismatched test cepstrum data onto the baseline cepstrum domain. Results for 2 different state-of-the-art codecs and a large number of different model topologies show encouraging improvement in performance compared to the mismatched cases.