Soft computing design for speech-based diagnosis of the human throat

The speech wave passes through the glottis and the vocal tract and finally arrives at the lips. The shapes of these organs modify the sounds we hear. A "tube model" can represent the vocal tract, and the various cross-sections progressing from the glottis to the lips characterize differing sounds. Such cross-sections may be linked to patient state. This may be NORMAL or ABNORMAL, damaged by cancer or unfortunate surgery or just old age. The author is interested in using vowel sounds as a possible diagnostic tool. Cross-sections of the vocal tract model can be calculated from the constants of a linear prediction coder, LPC, trained with a suitable vowel. The paper shows that an artificial neural network version of the LPC gives lower error than the original which should make it more useful for calculating cross-sectional areas. Two stages of fuzzy variables are used to diagnose a patient into two levels, NORMAL or ABNORMAL. Further analysis reinforces the diagnosis by using adverbs LIKELY or UNLIKELY to end up with one of four categories