基于发音特性的摩擦音和塞擦音分类算法

摘要

提出了一种基于发音特性的摩擦音和塞擦音分类方法,该方法首先基于Seneff听觉谱提取一组描述音段能量分布特性和谱统计量的特征参数,刻画两者在发音过程上的差异,然后采用支持向量机模型实现摩擦音和塞擦音的分类.实验结果表明,其干净语音分类准确率可以达到90.08％,信噪比为5dB的语音分类准确率可达到80.4％,与传统的基于时频能量分布特征的摩擦音和塞擦音分类方法相比,较大地提高了低信噪比下的性能.%A fricative and affricate classification method based on articulatory characteristic was proposed According to this method, the speech segment energy distributions and spectrum statistical features were first got based upon Seneff's auditory spectrum, and the differences of them were well described Then fricative and affricate classification was a-chieved using the support vector machine model The experimental results show that the classification accuracy is 90. 08% for clean speech ,80. 4% for noisy speech with the SNR of 5dB. Compared with the traditional time-frequency energy distribution features based fricative and affricate classification methods, the proposed method gets great performance improvement under low SNR.