Research On Forensics Of Digital Speech Processing History

With the continuous development of digital speech editing software,digital speech can be easily edited and processed.Some criminals can also use these speech editing software to disguise the speech segments,which will affects the judicial forensics.Pitch shifting,filtering,and noise adding operations are several types of speech processing that are relatively easy to modify speech.At the same time,if several processes are combined,it will have a greater impact on the forensics work.Therefore,this paper studies the mechanism of pitch shifting,low-pass filtering,high-pass filtering,noise adding processing and MP3 compression,as well as the processing chain of several processing combinations,and proposes corresponding detection algorithms.The research work of this thesis can be divided into the following three areas.Firstly,the mechanism of typical digital speech processing and the construction of the processing history dataset.This paper studies the mechanism of the four typical digital speech processing and MP3 compression,and build a dataset processing chain through the corresponding software.Secondly,a digital speech forensics algorithm suitable for a variety of processing trace is studied.By analyzing the time-frequency characteristics of the speech processed through four kinds of operations,it is proved that the four processing operations have a significant impact on the speech.Then,four processing operations are detected by the Mel Frequency Cepstral Coefficient(MFCC)statistical moment feature and the SVM combined classifier.The experimental results show that the method can effectively identify the original speech and the speech processed by four kinds of processing.Finally,digital speech processing chain forensics based on convolutional neural network.The MFCC statistical moment feature and the SVM combined voting classifier method have high algorithm complexi ty.For the digital speech processing chain,this paper proposes a convolutional neural network consisting of five special convolutional layers using the speech samples of the residual filter as input.By comparing the classification effects of convolutional neural networks with different network structures,a processing chain capable of effectively detecting a combination of four operations is obtained.At the same time,for the processing chain of MP3 compression and pitch shifting combined processing,convolutional neural network is used for feature extraction and classification.The experimental results show that the detection accuracy of the MP3 compression and pitch shifting combined processing can reach 97.57 %,and the overall detection rate is above 90%.