Quality improvements of zero-concatenation-cost chain based unit selection

Abstract

In our previous work, we introduced a zero-concatenation-cost (ZCC) chain based framework of unit-selection speech synthesis. This framework proved to be very fast as it reduced the computational load of a unit-selection system up to hundreds of time. Since the ZCC chain based algorithm principally prefers to select longer segments of speech, an increased number of audible artifacts were expected to occur at concatenation points of longer ZCC chains. Indeed, listening tests revealed a number of artifacts present in synthetic speech; however, the artifacts occurred in a similar extent in synthetic speech produced by both ZCC chain based and standard Viterbi search algorithms. In this paper, we focus on the sources of the artifacts and we propose improvements of the synthetic speech quality within the ZCC algorithm. The quality and computational demands of the improved ZCC algorithm are compared to the unit-selection algorithm based on the standard Viterbi search.