Audio Copyright Protection based on Quaternion Filtering
Bingkui Sun, Xun Jin · 2023
In this paper, we propose a method to solve the difficulty of audio model training convergence, the large data demand and the large dimension of audio feature vector storage space. In this method, one-dimensional data of audio signal is converted into two-dimensional spectral data, and quaternion Gabor filtering is used to suppress the background information of spectral graph to reduce the interference of data. In addition, the method also uses the window length and frame shift of different scales to capture the relationship between different objects, and uses the depth hash module to map the high-dimensional feature vector to the low-dimensional feature vector. We also introduce a probability function to make the learned sample more in line with the global distribution. In the experimental evaluation, the proposed method improves the performance of ESC dataset by 1% and GTZAN dataset by 1.3%.