Improving Phone Label Alignment Accuracy by Utilizing Voicing Information

The perceived quality of a concatenative speech synthesizer de-pends on an accurate alignment of unit labels and the corre-sponding waveform. This paper proposes to improve an HMM-based phone label alignment system by utilizing voicing infor-mation. Two different approaches have been investigated: (1) in-corporating voicing information into the feature vector and (2) acoustic post-processing based on changes in the estimated voic-ing state. Experiments on a Norwegian speech corpus show that both methods may improve alignment accuracy and that com-bined use may be beneficial. Informal listening tests indicate that the methods can yield improved quality for unit selection synthesis. 1