Segmental Intelligibility and Speech Interference Thresholds of High-Quality Synthetic Speech in Presence of Noise

Abstract

Technological advancement in the area of synthetic speech has made it increasingly difficult to distinguish quality of speech based solely on intelligibility scores obtained in benign laboratory conditions. Intelligibility scores obtained for natural speech and a high-quality text-to-speech system (DECtalk) are not substantially different. This study examined the perceived intelligibility and speech interference thresholds of DECtalk male and female voices and compared them with data obtained for natural speech. Results revealed that decreasing signal-to-noise levels had more deleterious effects on the perception of DECtalk male and female voices than on the perception of natural speech. Analysis of pattern of phoneme errors revealed that similar general patterns of errors tended to occur in DECtalk and in natural speech. The speech interference test did not demonstrate any significant difference between the DECtalk male and female voices. These results were supported by the absence of a significant difference between DECtalk male and female voices during intelligibility testing at different signal-to-noise ratios.