基于熵和支持向量机的病态嗓音识别

为了更好地分析实际短数据带噪的病态嗓音信号，利用近年来提出的样本熵、多尺度熵、模糊熵和分层熵的方法来提取嗓音的熵特征参数，并借鉴分层分解方法，提出分层多尺度熵和分层模糊熵，分别对测试集39例正常嗓音和36例病态嗓音进行支持向量机（SVM）识别。实验结果表明：三层分层熵、分层多尺度熵、分层模糊熵的识别率和稳定性均较分层前有提高。在耗时较短的情况下，提取2 000点病理嗓音数据的6种熵特征都能达到较好且较稳定的识别率。提取2 000点病理嗓音数据的三层分层模糊熵特征，能得到较好且较稳定的SVM识别率97.33%，较分层前的模糊熵特征识别率提高约4.00%。熵分析方法可推进病态嗓音研究向临床的应用，为临床分析诊断实时、短数据的带噪病理嗓音提供一定的参考。

Recognition of Pathological Voice Based on Entropy and Support Vector Machine

To solve the problems of short data and noisy recordings in pathological voice signals, this paper extracted some entropy feature parameters of pathological voice proposed in recent years, including sample entropy, multiscale entropy, fuzzy entropy and hierarchical entropy. Based on hierarchical decomposition method, we developed hierarchical multiscale entropy and hierarchical fuzzy entropy. Support vector machine (SVM) was used to distinguish the test set including 39 cases of normal and 36 cases of pathological voices. Results showed that three level hierarchical entropy，hierarchical multiscale entropy and hierarchical fuzzy entropy all achieved higher recognition rates and better stabilities using the proposed method. Pathological voice’s three level hierarchical fuzzy entropy feature got a better and more stable SVM recognition rate of 97.33% by extracting 2000 points. Compared with fuzzy entropy, the recognition rate was increased about 4.00%. The entropy method provide valuable preference for clinical analysis of short pathological voice time series contaminated by noise, which is benefit for clinical application of pathological voice analysis.