Audio Signal Forgery Detection Methods Based On Spectrogram And Pitch Synchronous

With the rapid development of digital audio technology, the digital audio signal is more and more widely applied to real life. As much powerful audio processing software are widely used, people who do not have professional skills can also manipulate the audio signal, such as cutting, stitching, copy and move. If the malicious tampered audio signal is used as court evidence or other important occasions, it will lead to serious social problems. Therefore, the importance of audio authentication technology is obvious.This paper focuses on audio signal forgery passive detection technology based on spectrogram and pitch synchronous, and gives different effective identification methods to detect audio signal upsampling and copy-move.(1) For audio signal upsampling detection, this paper presents two detection methods. The first methods base on the fact that spectrogram of audio signal with upsampling exists obvious mirror symmetry, while spectrogram of audio signal without upsampling is unique in each frequency band. First, we factorize each column of the normalized spectrogram into symmetric part and antisymmetric part. Then the proportion of the energy of the symmetric in the total energy is used to find the symmetry center and statistical method is used to determine whether the spectrogram exists a horizontal symmetry axis. Finally, the existence of the symmetry axis is used to determine whether the given audio signal suffered tamper. The second method analog the working of the auditory neuron cells by applying Non-negative matrix factorization spectrogram of audio signal, and then determine whether the audio signal is credible according activity of the auditory neuron cells.(2) For the audio signal copy-move detection, this paper presents a passive detection algorithm based on scale-invariant feature transform. The method first calculates the audio signal spectrogram and extracts all scale-invariant feature key points from the spectrogram. Then those key points were randomly divided into two subsets with same size. Best Bin First searching algorithm is used to find match key points between these two subsets, and for each subset recursive operation is used until all the matching key points is found. Finally, if a match is found, then the audio signal is considered has suffered copy-move forgery. And the copy-moved fragments of the audio signal are located by the area where the match key points gathered. (3) For the voice disguise detection, this paper presents a passive detection algorithm based on linear prediction coefficient. The method first frame the audio signal and extract linear prediction coefficient from each frame. Then vector quantization is applied to feature vector which is composed by linear prediction coefficient and its first order difference. Finally, the quantized feature vector is put in support vector machine to find out whether has suffered tamper.The computer simulation results show that the methods this paper proposed can effectively detect whether the audio signal has been tampered, and they all have better robustness.