Dysarthric speakers' intelligibility and speech characteristics in relation to computer speech recognition

Abstract

This research examined the use of the DragonDictate speech recognition system as a writing aid by 10 speakers with spastic dysarthria due to cerebral palsy with the ultimate goal of establishing prescription guidelines. The objectives were to (1) determine how intelligibility relates to recognition success; (2) determine those characteristics of speech, fluency, and voice that affect the accuracy level of the voice recognizer; and (3) investigate individual profiles of recognition over a series of dictations among high- and low-intelligibility dysarthric speakers versus nondisabled controls. Correlations between intelligibility measures and recognition success measures were strong. The total number of voice features correlated moderately with the number of readings needed to reach 80% recognition levels and with variability in recognition. The frequency of pauses showed a strong positive association with the number of readings subjects needed to reach 80% recognition. The frequency of nonspeech sounds was also associated with the number of readings needed to reach 80% recognition levels. Dysfluencies did not appear to exert a strong effect on recognition in this group of subjects. Most low-intelligibility subjects showed considerable variability in percentage word recognition from dictation to dictation, with fatigue being an important factor. High-intelligibility subjects showed less variability across dictations, with final scores close to the nondisabled controls. Two subjects in the low-intelligibility group also achieved final scores in the normal range.