Increasing the Precision of Dysarthric Speech Intelligibility and Severity Level Estimate

Dysarthria is a speech disorder often characterized by slow speech with reduced intelligibility. Automated assessment of the severity-level and intelligibility of dysarthric speech can improve the efficiency and reliability of clinical assessment as well as benefit automatic speech recognition systems (ASR). However, in order to evaluate them, there are not sentence-level severity and intelligibility label. We only have access to speaker-per-level severity and intelligibility labels. This is a problem as dysarthric talkers might be able to produce some intelligible utterances due to frequent use and short utterances. Therefore, label based analysis might not be very accurate. To address this problem, we explore methods to estimate the severity-level and speech intelligibility in dysarthria given discrete speaker-level labeling in the training set. To accomplish this, we propose a machine learning based method using one-dimensional Convolutional Neural Networks (1-D CNN). The TORGO dataset is used to test the performance of the proposed method, with the UASpeech dataset used for Transfer learning (TL). To evaluate, an Averaged Ranking Score (ARS) and intelligibility probability distribution are used. Our findings demonstrate that the proposed method can assess speakers based on severity-level and intelligibility to provide a more granular analysis of factors underlying speech intelligibility deficits associated with dysarthria.