A Comparative Study of WavLM-Based Speech Representations and Handcrafted Acoustic Features

The growing quality of artificially created speeches has created concerns regarding the use of synthetic voices for impersonation and voice-based scams. This research investigates the use of machine learning to identify synthetic speeches and distinguish them from real speeches. It compares two different deepfake voice detection pipelines, the experiment used 17,870 approximately two-second audio clips from the Fake-or-Real (FoR) dataset. To reduce the risk of data leakage, the dataset was reorganized into source- independent training, validation sets and testing sets so that the clips from the original recording does not occur in multiple sets.The first pipeline used the WavLM model to generate 768 dimensional speech embeddings classified using logistic regression. The second detection pipeline extracted 36 handcrafted acoustic features and classified them using CatBoost. On clean test audio, the pipeline using CatBoost achieved 99.07% accuracy and an Equal Error Rate (EER) of 0.91% whereas the WavLM pipeline with 97.51% accuracy and 2.63% EER.Then Robustness was evaluated using the addition of simulated Gaussian noise at 20 dB and 10 dB signal to noise ratios and by applying 32 kbps MP3 compression. Although CatBoost performed well in clean audio, its performance went down substantially after the introduction of Gaussian noise. WavLM retained considerably stronger performance under these conditions. Both pipelines were relatively effective after MP3 compression. The results illustrate that stronger performance with clean audio does not necessarily indicate greater robustness to audio degradation.