<p>Parkinson’s disease (PD) is a progressive neurodegenerative disorder that affects both motor and non-motor functions, with speech impairment being one of its earliest and most prevalent symptoms. Detecting PD through voice analysis offers a non-invasive, accessible, and cost-effective approach for early diagnosis and disease monitoring. In this study, we propose a robust deep-learning framework for PD detection using acoustic features extracted from speech recordings. Utilizing the publicly available MDVR-KCL dataset, the recordings are first preprocessed through a voice detection algorithm to isolate relevant speech segments. These segments are then concatenated and divided into fixed 2-second windows from which 43 acoustic features are extracted, including Mel-frequency cepstral coefficients, pitch, formants, spectral, and harmonic descriptors. The resulting features serve as input to various recurrent neural network architectures, namely long short-term memory (LSTM), gated recurrent unit, and bidirectional LSTM (BiLSTM), augmented with attention mechanisms to enhance temporal focus. Experimental results show that the BiLSTM-Attention model outperforms all others, achieving an average classification accuracy of 99.85% over 5-fold cross-validation. Statistical analyses confirm the discriminative power of the extracted features. These findings highlight the effectiveness of combining intelligent preprocessing, acoustic signal analysis, and attention-based deep learning to support accurate and scalable voice-based PD diagnostics.</p>