A large-scale comparison of the intelligibility of unit-selection and deep-neural-network parametric synthetic voices generated from dysarthric speech

Abstract

We report on a large-scale intelligibility study of 773 listeners who transcribed semantically unpredictable stimuli generated from 4 synthetic voices: two synthetic voices, a unit-selection voice and one based on deep neural network (DNN) parametric synthesis, from the recordings of each of two dysarthric speakers. Intelligibity was calculated as the normalized phonemic edit distance (NPED) between perceived and actual transcriptions. The DNN-based voices (NPED 0.286) were significantly more intelligible (p < .001) than the unit-selection voices (NPED 0.316), although there was a significant interaction (p = .034) with the structure of the synthesized sentence. Counterintuitively, the voices generated from the more severely dysarthric speaker (NPED 0.272) were signficantly more intelligible (p < 0.001) than the other speaker's voices (NPED 0.329). Post-hoc analysis demonstrated that while the more dysarthric speaker's speech had poorer vocal quality, was measurably slower and more variable in duration, and had a smaller vowel space, this speaker also had significantly (p > .001) higher formant amplitudes, narrower bandwidths, and less frequency variance; she also had more careful articulation during recording. We conclude that individual differences (some within the speaker's control) can override gross measures of dysarthria in determining synthetic voice intelligibility.