摘要
提出了采用模糊有向图支持向量机(FDGSVM)对基于输出的多语言语音样本进行语音质量评价的一种新方法.将多个可进行两类分类的模糊支持向量机组织成具有惟一根节点的有向图结构,得到多类分类器FDGSVM;提取待测语音信号的Mel倒谱系数并将其作为特征向量,再通过FDGSVM将特征向量映射到非线性划分的主观平均意见评分(MOS)区间,映射值即为输出的语音质量的客观评价结果.实验结果表明,所提算法获得的评测结果与主观MOS评价之间的相关度,在闭集测试时可达0.91,在开集测试时可达0.88.
A novel approach to output-based speech quality evaluation using fuzzy directed graph support vector machine （FDGSVM） is proposed. Binary classifiers of fuzzy support vector machine are organized into a structure of directed graph with a unique root to form the multi-classifier FDGSVM. Mel cepstrum coefficients are extracted from multi-lingual speech samples and regarded as eigenvectors, and mapped to different non-linear partition of subjective mean opinion score （MOS） through FDGSVM. The mapped scores are the out-based speech quality evaluation results. Experimental results show that the correlation between the result of the proposed approach and MOS is up to 0. 91 in close-set test and 0.88 in open set test.
出处
《西安交通大学学报》
EI
CAS
CSCD
北大核心
2006年第2期199-202,共4页
Journal of Xi'an Jiaotong University
基金
国家高技术研究发展计划资助项目(2003AA148010)
关键词
模糊有向图
支持向量机
MEL倒谱系数
语音质量
客观评价
fuzzy directed graph
support vector machine
Mel-cepstrum coefficient
speech quality
objective evaluation
