Identifying Concatenation Discontinuities by Hierarchical Divisive Clustering of Pitch Contours
In this paper, we present the results of a clustering experiment, the aim of which was to show whether or not the proximity of pitch contours is sufficient condition for perceptually smooth transitions at concatenation points in concatenative speech synthesis. The experiment was motivated by a previous finding which had shown that the support vector machine (SVM) classifiers are capable of separating with a high accuracy perceptually continuous and discontinuous joins using the pitch contours extracted from the vicinity of concatenation points as predictors. The experiment has shown that clustering of observations in a form of pitch contours represented in different scales using the euclidean distance as a metric does not prove to be a reliable way of identifying discontinuities at concatenation points.
Keywordsspeech synthesis unit selection concatenation cost pitch contours hierarchical divisive clustering
Unable to display preview. Download preview PDF.
- 1.Hunt, A., Black, A.: Unit selection in a concatenative speech synthesis system using a large speech database. In: ICASSP 1996, Atlanta, Georgia, vol. 1, pp. 373–376 (1996)Google Scholar
- 4.Bellegarda, J.R.: A novel discontinuity metric for unit selection text–to–speech synthesis. In: SSW5 2004, Pittsburgh, PA, USA, pp. 133–138 (2004)Google Scholar
- 5.Vepa, J.: Join cost for unit selection speech synthesis. Ph.D. thesis, University of Edinburgh (2004)Google Scholar
- 9.Legát, M., Matoušek, J.: Analysis of data collected in listening tests for the purpose of evaluation of concatenation cost functions. In: Habernal, I., Matoušek, V. (eds.) TSD 2011. LNCS(LNAI), vol. 6836, pp. 33–40. Springer, Heidelberg (2011)Google Scholar