Presentation of problem: Performance of jitter in discriminating between normal and dysphonic speakers

The data we deal with are acoustic feature values computed from speech signals produced by normal control speakers and speakers suffering from a laryngeal pathology. The problem is (i) to compare the discriminatory performance of a given acoustic feature with other features of the same family; (ii) to compare the performance of corresponding features in studies carried out on corpora which were not identical and which did not contain the same number of speakers; (iii) to express quantitatively the share of laryngeal pathologies which could be detected reliably by acoustic means alone. We propose to accomplish tasks (i) to (iii) by separating pooled control and risk groups on the basis of the acoustic feature values alone and by evaluating the quality of approximation of the original groups so realized. In order to evaluate this quality we compute a merit factor satisfying the following properties: (1) to be independent of the number of speakers contained in the pooled control and risk groups; (2) to increase and decrease linearly with the detection and false alarm rates, respectively; (3) to take on values between 0 (no discrimination at all) and 1 (perfect discrimination). We also address the question of the confidence one may have in the merit factor values so achieved, i.e. we describe a method of computing the probability that a merit factor value might have been realized by chance alone.