A comparison of spectral continuity measures as a join cost in concatenative speech synthesis
Brian W Kirkpatrick, Darragh O’Brien, Ronan Scaife · 2006
The quality of concatenative speech synthesis depends on the cost function employed for unit selection. Cost functions for spectral continuity are difficult to define and standard measures used for this task often do not accurately reflect human perception of discontinuity across a concatenated join. We compare a set of standard distance measures for the task of detecting audible discontinuities and introduce a new measure. A perceptual experiment is described that was used to relate each measure to human perception of discontinuities. The impact of window length on feature extraction and subsequent detection of discontinuities is investigated. The distance measure approach to detecting audible discontinuities is extended to a feature space based representation and feature transformations are investigated as a means of improving discontinuity detection. Receiver operating characteristic (ROC) curves is used to compare the results, which indicate that the feature space approach improves on the performance of standard measures.