A Review of Automated Intelligibility Assessment for Dysarthric Speakers

Automated dysarthria intelligibility assessment offers the opportunity to develop reliable, low-cost, and scalable tools, which help to solve current shortcomings of manual and subjective intelligibility assessments. This paper reviews the literature regarding automated intelligibility assessment, identifying the highest performing published models and concluding on promising avenues for further research. Our review shows that most of the existing work were able to achieve very high accuracies. However, we have found that most of these studies validated their models using speech samples of the same speakers used in training, making their results less generalizable. Furthermore, there is a lack of study on how well these models perform on speakers from different datasets or different microphone setups. This lack of generalizability has implications to the real-life application of these models. Future research directions could include the use of more robust methods of validation such as using unseen speakers, as well as incorporating speakers from different datasets. This would provide confidence that the models are generalized and therefore allow them to be used in real-world clinical practice.