Cantonese verbal information verification system using GMM-based anti-model

Abstract

Verbal information verification (VIV) is one of the approaches for speaker authentication. It is a process in which the spoken utterance of a claimed speaker is verified against the key information in a speaker's registered profile. VIV in English has been extensively studied and there has also been some work on Mandarin VIV. In the paper, we study the VIV for users who speak Cantonese, the most commonly used dialect in Southern China and Hong Kong. We propose a new technique for anti-modeling. It uses context independent Gaussian mixture model (GMM) instead of the conventional hidden Markov model (HMM). Experiments on 50 Cantonese native speakers show that the proposed method provides better separation of verification scores of claimant utterances from that of imposter utterances than the HMM based method. An equal error rate of 0.00% is attained with robust interval up to 15%, which manifests an excellent performance.