An evaluation of automatic phoneme segmentation for concatenative speech synthesis

摘要

This paper studies the performance of automatic phoneme segmentation in three viewpoints: (1) temporal precision, (2) effects on segment selection, and (3) effects on the naturalness of synthetic speech. The absolute error of the phoneme beginning time for the best 90% and worst 10% were respectively 4.6ms and 25.9ms, which are comparable to discrepancies among human labelers. Our segment selection algorithm was found to have ability to eliminate waveform segments with large temporal errors, although not perfectly. As the result of a perception test in which naturalness was paircompared between synthetic speeches generated from hand4abeled data and auto-labeled data, it was found that the difference is marginal in practice although the latter is statistically inferior.