Voice cutting trace positioning method based on double-domain depth features and attention mechanism

Abstract

The invention discloses a voice cutting trace positioning method based on double-domain depth features and an attention mechanism, and the method comprises the steps: obtaining a voice signal, carrying out the decompression and framing processing of the voice signal, calculating the discrete cosine transform of each voice frame, and obtaining the discrete cosine transform coefficient of each voice frame; extracting a time domain depth feature of each voice frame through a feature extraction sub-network of the one-dimensional convolutional neural network; extracting a frequency domain depth feature of a discrete cosine transform domain of each voice frame through a feature extraction sub-network based on a capsule network; splicing and fusing the double-domain depth features of each voice frame through a feature fusion full-connection layer to obtain a fused depth feature of each voice frame; and through a positioning module based on a multi-head attention mechanism, capturing contextual information fused with depth features, and positioning adaptive filtering traces in the cut voice. Whether the cutting trace exists on a section of suspicious recording or not is rapidly detected, and the cutting area is positioned.

Description

technical field [0001] The invention relates to the technical field of audio forensics in multimedia information security, in particular to a voice clipping trace positioning method based on a dual-domain depth feature and an attention mechanism. Background technique [0002] In recent years, the emergence of digital audio editing software has enabled some users who do not have knowledge of speech signal processing to easily process digital audio, so digital audio editing software has attracted a large number of users. With these convenient audio editing software, criminals can easily delete some words in a recording, thereby generating fake voices with completely different semantics. These fake voices are very likely to be used as evidence in court, or spread among the general public, causing a series of security issues. [0003] In the prior art, there is no good method to detect whether there is a clipping trace on a suspicious recording, and to locate the cropped area. ...

AI Technical Summary

Problems solved by technology

[0003] In the prior art, there is no good way to detect whether there is a cropping trace on a suspicious recording, and to locate the cropping area

Images