Generating emphasis from neutral speech using hierarchical perturbation model by decision tree and support vector machine

In a computer-aided pronunciation training (CAPT) system, corrective feedback is desired to provide contrastive comparisons between user´s and canonical pronunciations. This paper presents a hierarchical perturbation model to generate emphasis for English by modifying acoustic features of neutral speech to highlight such important speech segments. Synthesis of emphasis needs to be realized hierarchically at word, syllable and phone layers. A two-pass decision tree is constructed to cluster acoustic variations between emphatic and neutral speeches. The questions for decision tree construction are designed according to the above layers. The questions related to word and syllable layers are used to construct the main tree and then the questions related to phone layer are used to expand the leaves of main tree (deriving a set of subtrees). Support vector machines (SVMs) are used to predict acoustic variations for all the leaves of main tree (at word and syllable layers) and sub-trees (at phone layer). Experiments indicate that the proposed hierarchical perturbation model can generate emphatic speech with high quality for both naturalness and emphasis.