Abstract:
The conventional methods used for assessment or evaluation of articulation disorders are either manual or semiautomatic and hence, time consuming. Thus, an effort was made through ARF project (SH/CDN/ARF-ELECT1/2017-18) to automate the articulation test in Malayalam using Automatic Speech Recognition (ASR) techniques (Abraham et al., 2021). During validation of this system, it was noticed that, for speech of children with hearing impairment with severe articulation disorder, the ASR generated labels are sometimes absurd, which adversely affects its accuracy. So, there is a need for refinement of this existing system to improve the correlation with perceptual evaluation scores in children with severe disorders also. 30 children (Group I) with severe degree of hearing impairment and 120 normal healthy children between 3 and 7 years of age who are native Malayalam speakers participated in the study. The words spoken by the Group II participants were recorded using PRAT and was utilized for building the automatic speech recognizer. Word spoken by Group I participants were recorded using ‘record’ tab of the developed system and were used for training and testing. The speech processing model of the existing system was re-designed by modifying the automatic speech recognizer and the system for automatic identification of articulation errors and severity determination. In the modified system, a robust language model was established using CMU language modelling technique. The way in which SODA error was assessed was refined in the modified system. Also, the process with which PCC was calculated was also refined. Due to this the accuracy of SODA error detection and PCC calculation of the modified system was improved. The developed prototype was subjected to user testing on a few selected participants of Group I. Based on the feedback, appropriate modifications were done. The final prototype was validated by comparing the performance of the modified system with the results of perceptual evaluation.
