一种基于SVM的多特征参数清浊音判决算法

摘要

The composed voice of low bit rate vocoders usually have occasionally hoarseness, out-of-tone speech, caused by the low veracity of voiced-unvoiced classification. To solve the problem, a new improved algorithm based on Support Vector Machine combined with several characteristic parameters is proposed. Experimental results show that the algorithm greatly reduces the voiced-unvoiced classification error rate, and enhances the articulation and spontaneousness of the composed voices. Use this method in SELP (sinuous excitation linear prediction) vocoder,compared with other method with same bit rate, it has higher PESQ-MOS score, which shows its advantage.%为解决低速率声码器合成语音中，由于语音帧清浊判决不够准确而造成的偶发性嘶哑、机器音较重及变调等问题，提出一种基于支持向量机（Support Vector Machine，SVM）并结合多种语音特征参数的清浊音判决优化算法。实验结果显示，该算法能够有效降低清浊音的误判率，进而使合成语音的清晰度和自然度得到改善。将本算法应用到正弦激励线性预测算法中，在与相同码率的其他算法的比较实验中，得到较高的PESQ－MOS分，显示出一定的优势。