Automatic speech intelligibility detection for speakers with speech impairments: the identification of significant speech features

Abstract

Selection of relevant features is important for discriminating speech in detection based ASR system, thus contributing to the improved performance of the detector. In the context of speech impairments, speech errors can be discriminated from regular speech by adopting the appropriate discriminative speech features with high discriminative ability between the impaired and the control group. However, identification of suitable discriminative speech features for error detection in impaired speech was not well investigated in the literature. Characteristics of impaired speech are grossly different from regular speech, thus making the existing speech features to be less effective in recognizing the impaired speech. To overcome this gap, the speech features of impaired speech based on the prosody, pronunciation and voice quality are analyzed for identifying the significant speech features which are related to the intelligibility deficits. In this research, we investigate the relations of speech impairments due to cerebral palsy, and hearing impairment with the prosody, pronunciation, and voice quality. Later, we identify the relationship of the speech features with the speech intelligibility classification and the significant speech features in improving the discriminative ability of an automatic speech intelligibility detection system. The findings showed that prosody, pronunciation and voice quality features are statistically significant speech features for improving the detection ability of impaired speeches. Voice quality is identified as the best speech features with more discriminative power in detecting speech intelligibility of impaired speech.