Detection of Speech Smoothing on Very Short Clips

摘要

Audio editing software can easily be used to manipulate digital speech for forgery. Smoothing on the tampered boundary is usually performed to eliminate the obvious traces of forgery after tampering. This presents a considerable challenge for the forensic detection of tampered speech because the smoothing model is unknown and the smoothing operation often modifies only several tens of samples with the editing software. In this paper, we propose to apply six filtering models to approximate the smoothing in audio editing software for training the classifier. We analyze the impact of filtering operations on speech signals, especially on differential signals. On the basis of the local variance of the differential signal, we design a simple and yet efficient feature set. Theoretical analysis and extensive experiments show that the proposed features are very effective in detecting several common filtering operations on very short speech clips. The experimental results also show that the proposed method can detect unknown smoothing performed by commonly used audio editing software, such as Cooledit and Adobe Audition. This highlights the promising potential of the proposed method for use as a forgery localization tool of digital speech signals in practical forensic applications. The proposed method is capable of detecting smoothing on very short speech clips containing only several tens of samples and practical forgery used in audio editing software.