Detecting Forged Audio Files Using "Mixed Paste" Command: A Deep Learning Approach Based on Korean Phonemic Features.

Abstract

The ubiquity of smartphones today enables the widespread utilization of voice recording for diverse purposes. Consequently, the submission of voice recordings as digital evidence in legal proceedings has notably increased, alongside a rise in allegations of recording file forgery. This trend highlights the growing significance of audio file authentication. This study aims to develop a deep learning methodology capable of identifying forged files, particularly those altered using "Mixed Paste" commands, a technique not previously addressed. The proposed deep learning framework is a composite model, integrating a convolutional neural network and a long short-term memory model. It is designed based on the extraction of features from spectrograms and sequences of Korean consonant types. The training of this model utilizes an authentic dataset of forged audio recordings created on an iPhone, modified via "Mixed Paste", and encoded. This hybrid model demonstrates a high accuracy rate of 97.5%. To validate the model's efficacy, tests were conducted using various manipulated audio files. The findings reveal that the model's effectiveness is not contingent on the smartphone model or the audio editing software employed. We anticipate that this research will advance the field of audio forensics through a novel hybrid model approach.