A large-scale comparison of two voice synthesis techniques on intelligibility, naturalness, preferences, and attitudes toward voices banked by individuals with amyotrophic lateral sclerosis

Abstract

Amyotrophic lateral sclerosis (ALS) commonly results in the inability to produce natural speech, making speech-generating devices (SGDs) important. Historically, synthetic voices generated by SGDs were neither unique, nor age- or dialect-appropriate, which depersonalized SGD use. Voices generated by SGDs can now be customized via voice banking and should ideally sound uniquely like the individual’s natural speech, be intelligible, and elicit positive reactions from communication partners. This large-scale 2 x 2 mixed between- and within-participants design examined perceptions of 831 adult listeners regarding custom synthetic voices created for two individuals diagnosed with ALS via two synthesis systems in common clinical use (waveform concatenation and statistical parametric synthesis). The study explored relationships among synthesis system, dysarthria severity, synthetic speech intelligibility, naturalness, and preferences, and also provided a preliminary examination of attitudes regarding the custom synthetic voices. Synthetic voices generated via statistical parametric synthesis trained on deep neural networks were more intelligible, natural, and preferred than voices produced via waveform concatenation, and were associated with more positive attitudes. The custom synthetic voice created from moderately dysarthric speech was more intelligible than the voice created from mildly dysarthric speech. Clinical implications and factors that may have contributed to the relative intelligibilities are discussed.