Adapting Dynamic Time Warping to the Speech of the Hearing Impaired

Abstract

One service provided by our application 'Speech Assistant System' assisting the teaching of the hearing impaired to speak is the automatic assessment of words and sentences in the course of practice and feedback to the person. Individual speech sounds can only be correctly evaluated if they are compared with the appropriate reference speech sounds. This requires segmenting the speech to be examined. The methods currently known do not give sufficiently correct results for the speech of the hearing impaired, which is often so distorted and halting so that it prevents understanding. The paper presents a reference generation method suitable for segmenting distorted speech, a modification of dynamic time warping and its comparison with traditional methods. The procedure presented has been successfully used for the automatic assessment of the pronunciation of the hearing impaired.