COMPARISON OF HIDDEN MARKOV MODEL TECHNIQUES FOR AUTOMATIC SPEAKER VERIFICATION IN REAL-WORLD CONDITIONS

In this paper, we compare two alternative approaches for speaker verification based on hidden Markov model (HMM) technology: single Gaussian HMMs and different types of tied multi-Gaussian HMMs. In order to assess the performance under real-world constraints, we tested each system using a database of connected digit strings recorded over local and long-distance telephone lines. According to our experiments, tied-mixture models were able to perform better than the single Gaussian approach provided that sufficient training data were available. However, our experiments indicate that the single Gaussian HMM approach is to be preferred for real-world speaker verification when only limited amounts of training data are available. Results will be discussed for both text-dependent and text-independent speaker verification.