Deceptive Speech Detection Based on Multimodal Features and Attention Enhancement Method
Yadan Sun, Wenrui Li, Xunbing Shen · 2025
This paper presents a multimodal deceptive speech detection framework integrating MFCC, Chroma, and Mel features with a BiLSTM-GRU network and dynamic attention mechanism. To address class imbalance and feature limitations, noise augmentation and class-weighted loss are employed. Experiments on the Real-life Deception Detection 2016 dataset show that the proposed model achieves 90.2% accuracy and an F1-score of 0.91, outperforming traditional methods and deep models such as SVM and BiLSTM. The framework demonstrates strong generalization and is well-suited for practical applications in voice authentication and forensic analysis.