Automatic detection of stridence in speech using the auditory model

• Stridence is appearance of intense and sharp whistling in speech.
• An algorithm for stridence detection using Patterson's auditory model is presented.
• Three levels of decision are applied according to the categorical perception.
• Automatic detection is similar to that obtained by trained speech therapist.

Stridence as a form of speech disorder in Serbian language is manifested by the appearance of an intense and sharp whistling. Its acoustic characteristics significantly affect the quality of verbal communication. Although various forms of stridence manifestation are successfully diagnosed by speech therapists, there is a need for the automatic detection and evaluation of stridence. In this paper, an algorithm for stridence detection using Patterson's auditory model is presented. The algorithm consists of three processing stages. In the first stage spectral analysis and masking effects are applied using Paterson's auditory model. In the second stage a contour of spectral peaks that best fits characteristic features of the stridence is selected in the time-frequency (TF) representation of the signal obtained by Patterson's auditory model. In the third stage hypothesis testing is performed with three decisions: D0 – no stridence, D1 – stridence, and D2 – unable to decide. The reliability of stridence detection is tested on the speech corpus of 16 speakers without stridence (with correct speech), 16 speakers without stridence but with some other speech sound disorders, and 16 speakers with stridence. Test results show high correspondence of subjective measures and automatic detection.