基于SVM的说话声实时检测

Real-time Voice Detection Based on SVM

研究了说话声音频数据的音频特征,并且利用SVM实现了说话声的实时检测.首先,对原始音频数据做预处理,然后对音频数据进行了3层小波分解,并提取了第3层低频系数的MFCC特征,同时提取了音频数据的质心、谱熵特征;其次,计算MFCC特征的均值、质心和谱熵的方差,由此构造了特征向量集;最后,利用SVM模型训练音频样本,并对测试集中的样本进行了测试和实时检测.实验表明,所提取的音频特征有效、合理,并且表现出良好的分类与检测性能.

The voice features are researched,and real-time detection of voice is achieved by using SVM.Firstly,in the stage of preprocessing for the audio data,the MFCC of 3rd layer low frequency coefficients were extracted after performing 3-layer wavelet decomposition,furthermore,the centroid and spectral entropy were also extracted for audio data.Secondly,the features vector set was constructed,including the mean MFCC,the variance of centroid and spectral entropy.Finally,the audio data were trained and tested by making use of SVM for implementing real-time voice detection.The experimental results show that the extracted audio features are effective and reasonable,and reprensent the satisfactory classification and detection performance.