Abstract: Voice forgery is detected by digital audio watermarking and pattern recovery techniques. The three methods used by digital watermarking mode for speech signal forgery detection are replacement, insertion and deletion. If some changes are made to the speech signal, the watermark mode will also change, and the mode recovery can be used to measure or detect whether the speech signal has been modified or forged. The proposed method uses the cyclic pattern embedding to overcome synchronizing problems of previous detection techniques. In addition, pattern recovery enhances the robustness to compression. This method has been tested and verified using six recording devices, which was used for collecting verbal data. The speech signals were sampled at the rate of 8 kHz and digitized at 16 bits resolution. Randomly chosen regions were substituted, removed, and compressed in MP3 at the rate of 16 kbps as well as in CELP at the rate of 11.5 kbps. The experiment shows the perfect detection for three kinds of forgeries and it proved the validity of the proposed method.

Key words: Speech authentication, Digital audio watermarking, Spread spectrum, Speech forgery

摘要： 通过数字音频水印和模式恢复技术检测语音伪造，数字水印模式已经用于语音信号伪造检测的三种方法为替换、插入和删除，如果在语音信号上做一些改变，则水印模式也会变化，通过模式恢复可以检测语音信号是否被修改和伪造。提出利用嵌入的循环模式的方法来解决检测技术的同步问题，此外，模式恢复对语音信号压缩增强鲁棒性，该方法使用6个用于收集语音数据的录音设备进行测试和验证，语音信号的采样速率为8 kHz，将语音信号数字化的分辨率设为16 bit，对速率为16 kbps的MP3和速率为11.5 bps的CELP中语音信号随机选择区域进行替换、删除和压缩处理。结果表明，三种方法的语音伪造信号完全可以检测到，证明了该方法的有效性。

关键词: 语音认证, 　数字音频水印, 　扩频, 　语音伪造