As deepfake attacks pose an increasing threat to biometric voice security, developing effective countermeasures is essential. This study presents a novel approach to detect audio deepfakes by utilizing the Wave2Vec2 framework. A customized post-processing technique was incorporated during inferencing on the In-the-Wild dataset. Additionally, five layers “dropout, dense, tanh, dropout, and output projection” were integrated into the Wave2Vec2 architecture, leading to enhanced classification accuracy that is on par with peer models. To further ensure the reliability of predictions, a unique algorithm was developed, which compares the model’s logits against a set threshold and iteratively repeats the process until a confident outcome is obtained. The fine-tuned Wave2Vec2 model demonstrated remarkable results in both configurations: the Grounded Wav2Vec2, which fine-tunes only the custom classification layers, achieved an accuracy of 99.42% and an equal error rate of 1.53%, while the Adaptive Wav2Vec2, which fine-tunes the entire model, reached an accuracy of 99.59% and an equal error rate of 1.4%. These findings align closely with state-of-the-art results for deepfake detection on similar datasets.