Speaker
Description
Motor speech disorders can often lead to reduced intelligibility, communication difficulties, ultimately affecting the person’s quality of life. In recent years, several initiatives have been launched to improve the processing of disordered speech [1]. However, these initiatives differ in their methodology and their degree of success.We are working on a collaborative project called the ‘Dysarthria’ project at the University of Szeged with two main goals in mind. First, we conduct detailed analysis of dysarthric speech while considering etiological aspects [2, 3]. Secondly, we aim to find solutions which intend to improve the intelligibility of dysarthric speech using automatic procedures [4, 5]. This study focuses on the intelligibility of Hungarian dysarthric speakers, comparing four perceptual metrics: speech intelligibility, speech severity, speech naturalness, and listener effort. We invited trained and untrained listeners in speech analysis to rate the level of intelligibility of dysarthric speech samples. We compared their ratings with the error rates of automatic speech recognition (ASR) systems. Previously [5], we discussed automatic speech evaluation methods using four ASR systems. We selected 15 dysarthric speech samples from the Hungarian Dysarthria Database [6]. Our dataset contained recordings from 7 mild, 5 moderate, and 3 severe dysarthric speakers, and 5 healthy controls [7]. Speech samples were evaluated based on the four metrics using a visual analog scale. The judgements of human listeners were compared with the error rates of ASR systems. Findings showed great variation in the evaluation of severity, the degree of severity correlates with intelligibility in both listener groups.