摘要：Prosodic phrase prediction is important for both the naturalness and intelligibility of Text-to-Speech (TTS) systems. To automatically generate feature templates of prosodic phrasing models, this paper proposes a hybrid approach which converts the rules generated by classification and regression tree (CART) into templates of transformation-based learning (TBL), and designs a hierarchical clustering based feature combination algorithm for maximum entropy (ME) model. While minimizing human supervision, TBL templates automatically generated by CART can provide good alternatives or beneficial supplement to manually summarized templates, and ME templates automatically generated by the proposed feature combination algorithm not only make an improvement of 3.1% on F-measure over manual templates, but also reduce the size of ME model by up to 79.0%.
关键词：prosodic phrase prediction;feature template generation;keyword selection;classification and regression tree (CART);transformation-based learning (TBL);maximum entropy (ME)
