Experimental evaluation of information conveyed through voiced, unvoiced, and transition segments of speech
K. Ganesan, Wen C. Lin · The Journal of the Acoustical Society of America · 1975
Studies in the perception of speech have so far been only of a qualitative nature using psychoacoustic experiments. In this paper a quantitative method of making statistical measures is described. The technique uses the concepts of Shannon's information theory and a measure of information content is formulated. An algorithm based on statistical pattern recognition techniques is used to segment speech into four different segment classes [K. Ganesan and W. C. Lin, “Statistical pattern recognition approach to speech segmentation,” J. Acoust. Soc. Am. 56, S31 (A) (1974)]. Spectrograms of each type of segment are then obtained through which the corresponding probability density functions are estimated. Let Cv, Ct, and Cu be the voiced, transition, and unvoiced segment classes respectively. Information conveyed through Cv is given by I(X|Cv) = ∑ x1,x2,x3 p(X) log p(X) − ∑ x1,x2,x3 p(X,Cv) log p(X|Cv), where X = {x1,x2,x3}, a vector point in the spectral space. Applying this formulation to all the three classes, we found that the information conveyed through Cv is significantly lower than that through Ct and Cu, which are approximately equal.