Parkinson's disease (PD) is a neurodegenerative disorder that affects more than 10 million people worldwide. Despite its prevalence, the detection of PD remains a complicated task, as no gold standard test has yet been developed to provide an accurate diagnosis. In this context, many recent studies have focused on the automatic detection and progression tracking of PD from voice-related characteristics, being feature engineering the most common approach. This work intends to address an existing research gap by introducing a novel strategy that analyzes raw voice waveforms. Despite recent advancements, one of the significant hurdles is still the lack of extensive and diverse datasets. This article also implements a data augmentation solution. Big Vocoder Slicing Adversarial Network (BigVSAN) is used to generate synthetic voice data that mimics the characteristics of real patients and healthy subjects. For the PD detection task, deep learning models such as ResNet, LSTM-FCN, InceptionTime, and CDIL-CNN are used. The experiments were performed using the speech task of sustained vowel /a/ in the PC-GITA database, which contains the recordings of healthy and PD subjects. CDIL-CNN achieves the best results, improving the accuracy by 15.87% (8.96%) compared to the model that does not use augmented data (from the best method found in the literature that uses voice waveforms). The results of this study indicate that models trained with raw waveforms showcase modest but promising performance, underlying the potential of audio analysis to improve the early detection of PD, providing a non-invasive and potentially remotely applicable method.