Acoustic screening for vocal pathology

A computer-based acoustic system for detecting signs of laryngeal pathology from the tape-recorded speech waveform (40 s of continuous speech) is described. The system uses a parallel processing technique for pitch detection, with phase compensation, low-pass filtering, nonlinear smoothing for intonational baseline measurement, and local perturbation analysis. This provides two types of data: intonational data derived from the smoothed fundamental frequency trend line and waveform perturbation, which reflects irregularity of vocal fold vibration. Acoustic results are presented for a control group (100 speakers) and compared with the results for 80 speakers with known laryngeal disorders. The value of these results in screening for the effect of laryngeal abnormalities is examined. Statistical treatment of the data allows over 80% of speakers with known pathology to be differentiated from the control group. A second stage of analysis examines the possibility of using acoustic profiles to discriminate between different classes of vocal fold pathology (e.g., carcinoma, nodules, and Reinke's oedema). The relative discriminatory power of individual and combined acoustic measures is discussed. [Work supported by British Medical Research Council.]