Digital audio tampering detection using ENF feature and LST-MInception net

Abstract:
In recent years, with the rapid development of audio editing technology, detecting the edited audio is becoming more and more difficult. This brings serious risks to personal information security, such as voice locking. Electronic network frequency (ENF) is a signal embedded in many audio recordings, and we can extract the relevant features and put them into neural networks for training. We combine LSTM network with inception network, which is widely used in audio processing, and add Inception element on this basis, so that network editing at different times can have a good detection effect. Experimental results show that this method can detect forged data in ENF speech database with 95.4% accuracy.