Improving diphone synthesis by adding context-sensitive diphones

One well-known problem with concatenative synthesis is the occurrence of audible discontinuities at concatenation points. Formant jumps across concatenation points suggest the problem is due to spectral differences. In a previous experiment (Klabbers & Yeldhuis, 1998), the results of a listening experiment were correlated with several spectral distance measures to find one that best predicts the audible discontinuities. The Kullback-Leibler distance proved to be the best measure. In this paper we demonstrate its use for clustering diphones with similar contexts. For each of the clusters, a limited number of context-sensitive diphones is added to the database to reduce the number of audible discontinuities. A new listening experiment was performed, which showed that a significant improvement can be obtained.