The concept of deepfakes has come to light with the exponential growth of the field of Generative Artificial Intelligence (AI). Moreover, social media, being the major entity of engagement, became the prime area for wrongdoers to abuse these deepfakes. Hence, the misuse of audio deepfakes is a global concern, comparable to the threats posed by other types of deepfakes, and Bangladesh is no exception. Most of these research works are found to be conducted using readily available English audio datasets. However, Bengali audio deepfake-related works and Bengali audio datasets consisting of real and deepfake speeches for research purposes are hardly available. In this study, we attempted to contribute to the field of deepfake audio detection involving the Bengali domain. We experimented with several traditional machine learning classification models, including Support Vector Machine (SVM), Multilayer Perceptron (MLP), Decision Tree, Random Forest, and AdaBoost. We created a primary dataset by collecting our own Bengali audio data, besides incorporated two secondary datasets, one entitled “Bangla Audio Dataset: Original and DeepFake Voices for AI-Based Voice Analysis and Detection” and the other “Bangla Common Voice Corpus”. We processed audio data and extracted seven distinct features for feature-based analysis of the models. We found the Mel-Frequency Cepstral Coefficient (MFCC) feature to be the most dominant among the selected features. Additionally, we performed a comparative analysis on the performance of the models; thus, SVM and MLP obtained the highest accuracy, 94.14 % and 92.66 % with MFCC, respectively.