In the age of Artificial Intelligence, the rise of AI-generated audios, also known as audio deepfakes, causes significant concerns, especially regarding the authenticity of voices. While some research has been conducted on AD audio deepfake detection, the focus has mainly been on languages like English, leaving a gap in the study of Bangla-specific deepfake detection. This paper pioneers the exploration of audio deepfake detection in Bangla by developing two models: a Recurrent Neural Network (RNN)-based Long Short-Term Memory (LSTM) using MFCC features and a custom WaveNet model using normalized raw audio files. The LSTM model achieved an accuracy of 89%, while WaveNet outperformed it with a 91% accuracy rate. This research faced several unique challenges. In the absence of a proper deepfake audio dataset, a custom dataset was made for this research work. Additionally, the phonetic and tonal complexity of the Bangla language posed difficulties in distinguishing subtle variations between authentic and synthetic speech. The diverse accents and dialects within Bangla further complicated model generalization, requiring robust preprocessing and feature extraction techniques. Despite these challenges, this research successfully identified differences between authentic and synthetic audio, significantly improving detection accuracy for the Bangla language. This work not only advances audio deepfake detection in Bangla but also underscores the need for linguistic diversity in technological innovation, paving the way for future studies in other under-represented languages.