Automated Scoring of the Oral Diadochokinesis Task: Accuracy of an Approach Using Sonority Line and Machine Learning

Verbal diadochokinesis (DDK) is a widely used clinical assessment of oral-motor function in movement disorders. Manual scoring of DDK is time-consuming, which prevents the assessment from being used at-scale in a remote context. We have developed a novel technique that can automatically score DDK rate based on the sonority-line method. We have managed to mitigate impacts such as consonant noise, syllable smearing, variable response rate and volume impact through using a machine-learning post-processing stage. 153 participants aged 18-65 completed automated DDK tests using the NeuroVocalix^TM platform. Participants repeated the syllables “pa”, “ta”, ”ka” (SMR), or the sequence “pa ta ka” (AMR) twice for 10 seconds each at their maximum rate. 700 DDK utterances were manually scored. We then analysed the signal and employed machine learning techniques during the post-processing stage. Overall mean rate estimation accuracy was 94.5%, with SMR showing an accuracy of 96% and AMR 89.3%. “Pa Ta Ka” repetitions (AMR) correspond to the largest errors and majority of anomalies. Our results suggest that automatically estimating syllable rate can be done accurately using the sonority line method although AMR can be less effectively estimated than SMR. This tool would contribute to reducing the need for clinicians to manually label each sample.