音色变换音频信号的篡改检测技术研究

Study on tamper detection technology for voice transformation

针对音色变换软件带来的社会安全问题,提出一种音频信号篡改检测方法。首先根据语音信号的混沌特性和人耳的听觉特性,利用美尔频率倒谱系数(Mel frequency cepstral coefficients,MFCC)特征提取原理,提取待测音频的杜芬频率倒谱系数(Duffing frequency cepstral coefficients,DFCC),然后将特征参数的幅度进行提升,利用支持向量机(SVM)将特征参数与语料库里的特征进行分类。分类成功的情况下,根据幅度提升的大小判断待测音频信号是否经过篡改;同时根据幅度提升的大小和待测音频的性别判断说话人的真实性别。大量的实验结果表明,该方法在音频信号的篡改检测和音频信号说话人的真实性别判断方面均具有较高的准确率,并且性能稳定。

To address the social security issue brought up by voice transformation software, a method for sound signal tamper detection is proposed. Firstly, with the extraction method of Mel frequency cepstral coefficients (MFCC),Duffing frequency cepstral coefficients(DFCC) characteristic parameters of audio signals are extracted based on the human hearing characteristics and chaos characteristics of speech signal. Then, the amplitude of characteristic parameters is enhanced and support vector machine (SVM) is used to classify the characteristic parameters and characteristics in corpus. In case of successful classification, the audio signal will be judged whether it is tampered as per the size of the amplitude enhanced. Meanwhile, the speaker gender will be judged according to the size of the amplitude enhanced and the gender of the audio. Through a large number of experiments, it shows that the method has stable performance and high accuracy both in the audio signal tampering detection and audio speaker real gender judgement.