基于MARS的语音清晰度客观评价  

Objective Evaluation Method for Speech Articulation Using Multivariate Adaptive Regression Splines

作　　者：沈刘平[1] 杨吉斌[1] 曹铁勇[1] 张雄伟[1] 孙新建 

机构地区：[1]解放军理工大学通信工程学院,南京210007 [2]南京军区73061部队,徐州221121

出　　处：《数据采集与处理》2008年第1期100-103,共4页Journal of Data Acquisition and Processing

摘　　要：提出了基于多元自适应回归样条法(Multivariate adaptive regression spline,MARS)的语音清晰度客观评价方法。该方法提取语音信号的Mel倒谱系数作为评估语音清晰度的候选特征参数。在Mel倒谱系数的失真距离基础上,利用MARS方法选出对语音清晰度影响较大的特征参数,并结合主观DRT分建立最佳客观预测模型,实现特征参数失真距离到客观DR∧T分的映射。仿真结果表明,分别采用训练集合样本和测试集合样本进行测试时,使用该方法评价的客观DR∧T分与主观DRT分的相关度,分别达到0.958和0.9102。An objective evaluation method for speech articulation based on the multivariate adaptive regression spline（MARS） is presented. Firstly, the Mel cepstral coefficients are extracted as candidate features for the articulation evaluation by the method. Then features are selected using MARS, and an objective forecast model is constructed to accomplish the mapping from the distortion measure diagnostic rhyme test （DRT） to the objective score. The simulation result shows that the correlation degree between objective and subjective DRT scores reaches 0. 958 when using training samples and 0. 910 2 when using testing samples. The evaluation results of the proposed method are obviously superior to that of other methods.

关 键 词：语音清晰度 客观评价 多元自适应回归样条法(MARS) MEL倒谱系数 

分 类 号：TN912.3[电子电信—通信与信息系统]