基于MFCC的混响效果识别研究  

Reverberation Level Recognition Based on MFCC

作　　者：马赛[1] 谢茜[1] 刘嘉胤 MA Sai;XIE Xi;LIU Jia-yin(Key Laboratory of Media Audio&Video,Ministry of Education,Communication University of China,Beijing 100024,China;Information Center of Shandong Tobacco Company,Jinan 250101,China)

机构地区：[1]中国传媒大学媒介音视频教育部重点实验室,北京100024 [2]山东省烟草公司信息中心,济南250101

出　　处：《中国传媒大学学报（自然科学版）》2017年第4期18-23,共6页Journal of Communication University of China：Science and Technology

摘　　要：直达声/混响声能量比(Direct-to-Reverberant Ratio,DRR)和混响时间(Reverberation Time,RT)是判断混响效果的两个重要参数。针对合成有声语音及元音EH,在给定的混响时间下,提取不同直达声/混响声能量比的混响语音信号的MFCC(Me-Frequency Cepstral Coefficients)特征,分别对其进行混响效果识别的10折交叉验证。利用高斯混合模型对训练集进行聚类分析,得到其概率分布函数,通过马氏距离(Mahalanobis Distance,MD)计算测试集的每个样本的混响效果概率,进而判断其混响效果等级。实验结果证明,合成有声语音基于MFCC的混响效果识别准确率可以达到90%以上,元音EH可以达到80%以上。Direct-to-Reverberant energy Ratio(DRR)and Reverberation Time(RT)are the primary parameters for reverberation strength judgement.Given some selected RT,cluster reverberant synthesized voiced speech and vowel EH at different DRR based on MFCC,and use 10-fold cross validation for reverberation level recognition,respectively.Train the training dataset by Gaussian Mixture Model to obtain the probability distribution,and calculate the test dataset probability via Mahalanobis Distance in order to achieve the recognition purpose.Experiments show that reverberant voiced speech recognition accuracy is higher than 90%,and reverberant vowel EH is higher than 80%.

关 键 词：混响 MFCC 交叉验证 高斯混合模型 

分 类 号：TN912.3[电子电信—通信与信息系统]