Multi-scale and multi-task learning for human audio forensics based on convolutional networks

Abstract

Audio forgery has become a growing challenge with rising social media usage. To reduce complexity and enable accurate recognition, this paper proposes an EfficientNet model with multi-scale spectral image features and multi-task learning for two scenarios: audio forgery and speaker identification. Audio signals are enhanced and expanded by multiscale audio segmentation, then different feature extraction techniques are used to transform them into spectral image features. Image recognition is performed by EfficientNet, and the two-stage voting method is used for multi-task learning. Experiments on Ali Tianchi datasets 1 & 2 show optimal results under Spectrogram-based spectral image features. Compared to multi-scale ResNet-152, Inception-v3, and Inception-Resnet-v2 models, the accuracy of the binary classification task was improved by 8.26%, 9.64%, and 2.46%, respectively (dataset 2, 12.80%, 11.60%, and 2.44%); the accuracy of the multi-classification task was improved by 9.61%, 8.56%, and 1.97%, respectively (dataset 2, 12.80%, 11.60%, and 2.44%).