Automatic Speech Intelligibility Testing for Thai Oral Surgical Patients
N. Pitaksirianant, Kwanchiva Saykhum, Chai Wutiwiwatchai, A. Chotimongko, Klong Nueng · 2011
This paper introduces the use of automatic speech recognition (ASR) in speech intelligibility testing for oral surgical patients caused by cancer, tumor, or fracture. The proposed automatic system aims to help reduce the effort and cost required by the conventional human listening test. We developed a phone based ASR system which outputs the best matched word out of a set of target words given a patient's recorded speech. To enhance the performance of the ASR system, an acoustic model is adapted toward the characteristics of each patient's voice. With an unsupervised speaker adaptation technique, the correlation between speech intelligibility scores given by the ASR system and the ones given by human judges was improved by 57% relatively. Although the overall correlation is still not sufficient for practical use at this state of the research, there is a possibility to apply the system for some specific types of patients such as fracture which the system achieved the correlation as high as 0.91. transcribe the utterances. The transcriptions can hence be used in the acoustic model adaptation process. Overall architectures of the baseline and the enhanced systems are given in Section III. Section IV describes experimental data taken from volunteer patients and experimental results in terms of statistical correlation between speech intelligibility scores given by the ASR system and the ones given by human judges. Result regarding each type of patient as well as the overall result are reported and discussed. Finally, a conclusion