Skip to Main Content
In conventional HMM-based speech synthesis framework, spectral features are modeled in one stream, and stream-dependent tree-based clustering was then applied for tying the model parameters. In this paper, we investigate several different stream-dependent tying structures for spectral features by splitting the feature vector into several streams. One splitting approach is to split each feature dimension into each stream. Another one is to split the static and dynamic features into different streams. Although splitting spectral features into different streams would ignore the correlation of context dependency between them, the number of model parameters can be optimized for each stream after stream-dependent clustering. From the experimental results, both splitting approaches can improve the quality of synthesized speech. However, the quality of synthesized speech became worse when we combined these two splitting approaches.