Speaker Verification System Using a Hierarchical Adaptive Network-Based Fuzzy Inference Systems (HANFIS)

We propose the use of a hierarchical adaptive network-based fuzzy inference system (HANFIS) for automated speaker verification of Persian speakers from their English pronunciation of words. The proposed method uses three classes of sound properties consisting of linear prediction coefficients (LPC), word timelength, intensity and pitch, as well as frequency properties from FFT analysis. Actual audio data is collected from fourteen Persian speakers who spoke English. False acceptance ratio and false rejection ratio as are evaluated for various HANFIS trained with different radius. Results indicate that vowel sounds can be a good indicator for more accurate speaker verification. Finally, the hierarchical architecture is shown to considerably improve performance than ANFIS.