Automatic language recognition based on discriminating features in pitch contours

A study of linguistic literature proposed that prosodic features in the pitch contours of languages could be the basis of an efficient automatic language recognition (ALR) system. In order to extract the pitch contour from speech signals, an accurate event detection pitch detector has been introduced. The latter is based on a time-scale representation, the dyadic wavelet transform (D_y WT), which aims at extracting the transients in the signal, based on the dilation of the analysis window. The proposed pitch detector is suitable for both low pitched and high pitched speakers, for non-stationary pitch periods and it is robust to noise. A set of feasible, discriminating features were extracted from the pitch contour and were used in a "k-nearest neighbor" classification technique to classify three languages. Results indicated an excellent distinction between a tone and a stress language, Xhosa and Afrikaans