Automatic error detection in alignments for speech synthesis

Abstract

The phonetic segmentation of recorded speech is a crucial factor in the quality of concatenative systems for speech synthesis. The authors describe a likelihood-based error detection process that can be used to flag possible errors in such a segmentation, with a view towards manual correction. It is shown that this process can be used to assist in the creation of high-accuracy segmentations. In particular, for an isiZulu corpus used in the creation of a unit-selection synthesizer, almost half of the errors that existed in a manual segmentation were detected by this process, while flagging less than a quarter of all segments. Different phoneme classes are handled with differing amounts of success, with vowels being the most troublesome