Estratégias de ampliação e balanceamento de dados para a detecção de áudios verdadeiros e falsos

Abstract

This work investigates the impact of data augmentation and class balancing strategies on improving the performance of supervised classification models in the detection of speech audio manipulated by text-to-speech (TTS) systems. The Fake or Real (FoR) dataset, developed by the Lassonde School of Engineering, was used. It consists of authentic and synthetic speech samples generated by different TTS architectures. Five data augmentation techniques were applied: noise addition, speed variation, pitch shifting, time-frequency masking, and random spectral equalization. To address class imbalance, Synthetic Minority Over-sampling Technique (SMOTE) and random undersampling were used. The audio signals were represented by Mel- Frequency Cepstral Coefficients (MFCCs), which served as input to the following classifiers: Random Forest (RF), Light Gradient Boosting Machine (LGBM), Naive Bayes, Long Short-Term Memory (LSTM) neural network, and k-Nearest Neighbors (KNN). The combination of data augmentation with balancing via SMOTE provided the best overall performance. In this scenario, the Random Forest (RF) model stood out, achieving accuracy and F1-score of 95%.