Segmental Intelligibility of Three Text-to-Speech Synthesis Methods in Reverberant Environments

Abstract

In this study, the segmental intelligibility of three currently available text-to-speech products under two reverberant conditions was investigated. The reverberation times used were 1.2 and 2.4 s simulating reverberation that may exist in a large room and a large hall with poor acoustics. The human speech had an overall intelligibility (whole words correct) of 95% and a phoneme error rate of 2.35% under reverberant conditions investigated in this study, which were not significantly different from those obtained in a nonreverberant controlled condition. In contrast, the overall intelligibility of text-to-speech voices was 68% and phoneme error rate was 14.48%, which indicated that that text-to-speech output suffers significantly in the same reverberant conditions. Implications of these findings for the improvement of text-to-speech products and the practice of AAC are discussed with suggestions for further research.