Audio tampering recognition algorithm based on improved neural network

Abstract

The invention discloses an audio tampering recognition algorithm based on an improved neural network. A spectrogram of any size is pooled into a CNNs structure represented by a spectrogram of a fixedlength and an LSTM structure with an attention mechanism; a MEL spectrogram and a frame level feature of a signal are introduced into a voice tampering recognition algorithm, and the frequency spectrum and the timing information of an audio signal are integrated; an improved pooling layer is added in the CNNs structure, so that the CNNs can input spectrogram of any size, and the problem that the audio length is not fixed is solved; the weight proportion of high-level features are excavated by virtue of the addition of the attention mechanism, so that the high-quality audio features are finallyobtained; an algorithm of decision fusion is carried out by utilizing the data fusion theory; and the recognition rate of the audio tampering recognition and the robustness of a model are improved. According to the method, the audio tampering can be effectively recognized, and the problem that the traditional audio tampering recognition rate is relatively low is overcome.

Description

technical field [0001] The invention belongs to the field of audio tampering, and in particular relates to an audio tampering recognition algorithm based on an improved neural network. Background technique [0002] The increasing maturity of digital audio editing technology has destroyed the authenticity and integrity of digital audio. When falsified audio is used as evidence in court, it will have a great impact on the judgment of the case. Therefore, judging whether the audio has been tampered with or not is an urgent problem to be solved by the relevant judicial departments. [0003] In 2005, Grigoras.C found that there was a power grid frequency component in the recording signal powered by mains power, and extracted the frequency characteristics of the power grid in the audio to be tested to match and compare with the data in the power grid frequency characteristic database of the power supply department. With a high degree of similarity, it is proposed for the first t...

AI Technical Summary

Problems solved by technology

Firstly, there is insufficient research on the characteristics of audio tampering recognition; secondly, the audio tampering recognition model, the existing audio tampering models are all traditional signal processing models, and machine learning and deep learning are rarely used for analysis

Images