Detecting speech authenticity has become increasingly critical to ensure trust, security, and integrity in modern digital communication systems. While significant progress has been made in AI-based voice detection for English and other high-resource languages, Bengali speech remains largely unexplored due to limited datasets, linguistic complexity, and resource constraints. To address this gap, we propose a Convolutional Neural Network (CNN)-based framework capable of distinguishing between human and AI-generated Bengali speech. The system leverages Mel-Frequency Cepstral Coefficients (MFCCs) for feature extraction, capturing key spectral properties of speech, and employs deep CNNs to learn robust and discriminative representations that enable accurate voice classification. Extensive experiments on a curated dataset demonstrate that our model achieves an F1-score above 98.5% under five-fold cross-validation, highlighting its strong generalization ability and reliability. This approach provides a practical and scalable solution for forensic audio analysis, mitigation of AI-generated misinformation, and secure voice-based authentication in Bengali-speaking applications. Furthermore, the methodology can be adapted to other low-resource languages, paving the way for more inclusive and globally applicable speech authenticity detection systems.