F0 feature extraction by polynomial regression function for monosyllabic Thai tone recognition

This paper presents a monosyllabic Thai tone recognition system. The system is composed of three main processes, fundamental frequency (F0) extraction from input speech signal, analysis of F0 contour for feature extraction, and classification of each tone using the extracted features. In the F0 feature extraction, the polynomial regression functions are employed to fit the segmented F0 curve where its coefficients are used as a feature vector. In tone recognition, we used the maximum a posteriori probability classifier (MAP) to classify a tone by assuming that the feature is a multidimensional Gaussian random variable. The hypothetical words used in this paper are composed of numerical words and monosyllabic Thai words. The vocabulary set is composed of the short vowel words, the long vowel words and have the effect of initial and final consonant on the shape of F0 contour. The experimental results show that by using the system as a speaker-dependent system, the maximum recognition rate is 96.20% using three-dimension feature vector. The speakerindependent recognition rates are 79.99% for male and 82.80% for female using four-dimension feature vector.