Cutting-edge progress in AI-based audio applications and speech technologies has improved the skillset of deepfake developers to alter the integrity of content and spread harmful falsehoods. Deepfakes are now capable of producing content that appears natural, rendering them tougher to recognize. The growing use of deepfake audio produces critical hindrances in cybersecurity, privacy protection, media trust, and safeguarding personal reputation. Technologies capable of detecting audio deepfakes are necessary to tackle the spread of deceptive practices. Sometimes it is incapable of determining the precise position of the synthetic segment in the audio, and the models fail to detect the stronger signal when subjected to fluctuations and distortions in the audio signal. In light of this, the study presents an innovative machine-learning approach to detect deepfake audio. The research is initiated by systematically curating the FOR-2SEC dataset, FOR-REREC dataset, two versions of the FoR (Fake or Real) dataset, and Bangla Audio Dataset, containing authentic and synthetic audio recording collections. In the presented framework, audio feature sets are retrieved from the dataset leveraging Mel-Frequency Cepstral Coefficients (MFCC). A precisely optimized Convolutional Neural Network (CNN) integrated with a Bidirectional Long Short-Term Memory (BiLSTM) network is introduced for classification tasks. Our approach exhibits remarkable effectiveness in differentiating authentic audio from deepfake recordings through systematic experimentation and optimization of model parameters. The framework attained a comprehensive test accuracy of 98.81% on the FOR-2SEC dataset, 97.28% on the FOR-REREC dataset, and 99.72% accuracy for the Bangla Audio Dataset.