Vowels' Articulatory Location Classification based on Formant Frequency

Modelling vowels is a crucial step in many applications such as speech understanding, human age analysis, accents differentiation, security and E-learning. This paper proposes building a statistical model for the US English vowels based on formant frequencies, which are extracted from real speech samples. Vowels are classified according to their articulatory-acoustic characteristics into front and back vowels. Three different approaches have been tested in this research, the K-means clustering, Fisher's linear discriminant analysis (LDA) and non-linear Gaussian mixture model (GMM). The proposed GMM model is used to classify vowels with the help of the Expectation Maximization (EM) algorithm. The results show that K-means analysis did not produce a convenient classification of vowel formants. Meanwhile, the automatic perception is highly achievable when formant data are projected on different coordinates using Fisher's LDA (FLDA). The classification error in the FLDA is close to that of the quadratic discriminant function of the EM algorithm. Although EM algorithm that was based on GMM gave better classification when it was tested, Fisher's LDA gave nearly the same results with significantly less computations.