Concatenation Artifact Detection Trained from Listeners Evaluations
Unit selection is known for its ability to produce high-quality synthetic speech. In contrast with HMM-based synthesis, it produces more natural speech but it may suffer from sudden quality drops at concatenation points. The danger of quality deterioration can be reduced (but, unfortunately, not eliminated) by using very large speech corpora. In this paper, our first experiment with automatic artifact detection is presented. Firstly, a brief description of artifacts is given. Then, a listening test experiment, in which listeners evaluated speech synthesis artifacts, is described. The data gathered during the listening test were then used to train an SVM classifier. Finally, results of the SVM-based artifact detection in synthetic speech are discussed.
Keywordsspeech synthesis unit selection error detection
Unable to display preview. Download preview PDF.
- 1.Klabbers, E., Veldhuis, R.: On the reduction of concatenation artefacts in diphone synthesis. In: Proc. ICSLP, Sidney, Australia, pp. 1983–1986 (1998)Google Scholar
- 2.Pantazis, Y., Stylianou, Y., Klabbers, E.: Discontinuity detection in concatenated speech synthesis based on nonlinear speech analysis. In: Proc. INTERSPEECH, Lisbon, Portugal, pp. 2817–2820 (2005)Google Scholar
- 3.Lu, H., Wei, S., Dai, L., Wang, R.H.: Automatic error detection for unit selection speech synthesis using log likelihood ratio based SVM classifier. In: Proc. INTERSPEECH, Makuhari, Japan, pp. 162–165 (2010)Google Scholar
- 5.Tihelka, D., Kala, J., Matoušek, J.: Enhancements of Viterbi search for fast unit selection synthesis. In: Proc. INTERSPEECH, Makuhari, Japan, pp. 174–177 (2010)Google Scholar