Comparison of text-independent speaker recognition methods on telephone speech with acoustic mismatch

We compare speaker recognition performance of vector quantization (VQ), Gaussian mixture modeling (GMM) and the Arithmetic Harmonic Sphericity measure (AHS) in adverse telephone speech conditions. The aim is to address the question: how do multimodal VQ and GMM typically compare to the simpler unimodal AHS for matched and mismatched training and testing environments? We study identification (closed set) and verification errors on a new multi environment database. We consider LPC and PLP features as well as their RASTA derivatives. We conclude that RASTA processing can remove redundancies from the features. We affirm that even when we use channel and noise compensation schemes, speaker recognition errors remain high when there is acoustic mismatch.