Preliminary results using inverse filtering of speech for automatic evaluation of laryngeal pathology Free

Inverse filtering of speech by an approximation of the supraglottal transfer characteristic produces a residue signal which appears to be a good indicator of laryngeal pathology [Koike and Markel, Ann. Oto. Rhinol. Laryngol. 84, 117 (1975)]. In the present study, residue signals from isolated occurrences of /a/ for 17 normal and 22 pathologic male and female speakers were used to compute the amplitude of the autocorrelation peak corresponding to the pitch period, the coefficient of excess, the pitch and the amplitude perturbation quotients, and the inverse filter and residue spectral flatness measures. A closed two‐choice Bayes maximum likelihood test based on these features resulted in correct identification of 100% of the normal speakers and 95% of the pathologic speakers. Respective discrimination in an open two‐choice test, where the speakers were divided into reference and test groups, was 72% and 91%. Results were tabulated for different pathologies, analysis conditions and data base sizes. These findings indicate the feasibility of using statistical methods to automatically discriminate between normal and pathologic voices. [This research supported by the National Institutes of Health under Grant No. NS01877.]