Audio manipulation detection based on wavelet spectrogram and multidimensional feature fusion with dual-channel CNN

With the prevalence of editing software, audio manipulation has become increasingly easy, posing a serious threat to the authenticity and integrity of audio data. Among various forms of manipulation, audio splicing and copy-move forgery are common. However, current audio forgery detection technologies typically can only detect one form of tampering, which presents significant limitations in practical applications. To address this issue, we propose a novel approach based on wavelet spectrograms and dual-channel convolutional neural network (DCCNN) for detecting audio splicing and copy-move forgery simultaneously. Initially, we transform the audio signals into wavelet spectrograms and then employ DCCNN for feature extraction and classification. Compared with the traditional convolutional neural network, this algorithm adopts a dual-channel convolutional structure combined with a multi-scale feature fusion technique, which is able to extract the local and global features of the wavelet spectrogram more effectively. By using convolution kernels of different sizes to perform convolution operations on the feature map and fusing these features, the feature expression ability of the image is enriched. Experiment result show that this method can simultaneously detect traces of splicing and copy-move forgery, and exhibits higher detection accuracy and robustness in audio forgery detection tasks. The algorithm was tested on dataset spliced and copy-move forged samples created from the Arabic speech corpus, TIMIT databases and ADD databsed, respectively. Subsequently, compared to existing state-of-the-art algorithms for audio splicing and copy-move forgery detection, the experimental results show that the proposed algorithm performs excellently in accuracy and significantly state-of-the-art methods.