VT-AMDF pitch detection algorithm and Thai tone recognition system

This paper, we evaluated VT-AMDF pitch detection algorithm with Thai tone recognition system. The varied time interval of average magnitude different function (VT-AMDF), the original AMDF, and the circular AMDF (CAMDF) are used to extract fundamental frequency (F₀) and compared recognition performance. F₀ data, extracted from those three pitch detection algorithms are separated and formed a contour and a subsection F₀ which used as tone features of Hidden Markov Models (HMM). Training and testing utterances were collected from 26 speakers uttering seven Thai isolated words in five difference tones. The result shows that the VT-AMDF algorithm yields more accurate tone recognition than the others when cooperated with minimum detection method or threshold detection method in 256 samples per frame condition.