An improved tone labeling and prediction method with non-uniform segmentation of F0 contour

Abstract

This paper proposes a tone labeling technique for tonal language speech synthesis. Non-uniform segmentation using Viterbi alignment is introduced to determine the boundaries to get F0 symbols, which are used as tonal label to eliminate the mismatch between tone patterns and F0 contours of training data. During context clustering, the tendency of adjacent F0 state distributions are captured by the state-based phonetic trees. Means of tone model states are directly quantized to get full tonal label in the synthesis stage. Both objective and subjective experiment results show that the proposed technique can improve the perceptual prosody of synthetic speech of non-professional speakers.