基于最小生成误差的HMM模型聚类自动优化

Minimum Generation Error Based Optimization of HMM Model Clustering for Speech Synthesis

为改善决策树聚类的效果,避免可能出现的聚类模型过训练或欠训练的情况,提出一种基于最小生成误差以及通过交叉验证优化最小描述距离(MDL)因子选取的方法.文中通过计算交叉验证中的生成误差选择MDL因子,从而优化决策树大小.实验结果表明,此方法相对传统的固定MDL门限设定方法,更有效提升合成语音的音质和自然度.

To improve the decision tree clustering and avoid possible clustered model over-training and less-training,a minimal generation error criterion and cross-validation（CV） based minimal description length factor optimizing method is introduced.CV based generation error is calculated to optimize the scale of the decision tree.Results of both subjective and objective tests show that synthesized speech by the proposed method outperforms the synthesized speech by the baseline one system in both quality and naturalness.