Compression Detection of Audio Waveforms Based on Stacked Autoencoders

摘要

With the easy acquisition of digital recordings, the field of audio foren-sics has become increasingly prominent. Detection of audio compression history is an important issue in the field of audio forensics. In this paper, a detection framework is proposed to detect whether a given audio waveform is an original waveform or a decompressed one. We extract the spectrum features from the frequency domain and then adopt a stacked autoencoder to effectively detect the frame-level audio fragments to distinguish between the original audio frames and the decompressed audio frames. Then, a majority voting algorithm is applied to make the final decision for an audio clip. Our analysis focuses on multi-time compressed audio, including single compression, double compression, triple compression and even four-time compression in three kinds of compression formats. The experimental results show that the proposed framework can effectively detect multi-time compressed audio. Furthermore, the proposed framework can also estimate the compression bitrate.