论文部分内容阅读
该文提出了一种基于Viterbi解码的中文合成音库韵律短语边界自动标注方法,以降低大语料库单元拼接合成系统的构建成本。该方法分为模型训练和韵律标注两阶段:模型训练阶段得到频谱、基频和音素时长的上下文相关隐Markov模型(hidden Markov model,HMM);标注阶段借助训练得到的模型采用Viterbi解码完成韵律短语自动标注。实验结果表明:该方法进行韵律短语边界标注时的F-score值达到77.64%,超过了人工标注时不同标注人员之间的一致性水平;另外该方法可以方便地增加待标注韵律属性,具有良好的扩展性。
In this paper, an automatic labeling method for prosodic phrase boundaries of Chinese synth bases based on Viterbi decoding is proposed to reduce the cost of constructing large corpus unit splicing system. This method is divided into two phases: model training and prosodic labeling. The hidden Markov model (HMM), which obtains the spectral, fundamental frequency and phoneme duration in the model training phase, completes the prosodic phrase by using Viterbi decoding in the marking phase Automatic labeling. The experimental results show that the F-score of the method is 77.64% when the prosodic phrase boundary is marked, which exceeds the consistency level of the different annotated personnel when the artificial annotation is performed. In addition, the method can conveniently increase the prosodic property to be annotated and has good Scalability.