Continuous monitoring is imperative for the early detection of Alzheimer's Disease (AD). Speech data is a promising modality for continuous monitoring of AD, as it can be collected in real time and evaluated for subtle changes that may indicate disease progression. This work presents comparison of different cepstral encoding techniques for spontaneous speech signals that can aid in Alzheimer’s disease classification using EfficientNet architecture. For this, spontaneous speech samples are obtained from publicly available dataset. These audio samples are preprocessed and spectrograms such as logmel, Mel-Frequency Cepstral Coefficients (MFCC), Gammatone Frequency Cepstral Coefficients (GFCC), and BFCC (Bark Frequency Cepstral Coefficients) are obtained. The spectrograms and their derivatives are given as input into a pre-trained EfficientNet-B0 model with a combination of fully connected networks. This model is evaluated using various performance metrics and cross-validated for five folds. The best-performing cepstral feature and its derivatives are selected for further hyperparameter tuning such as optimizer, batch size, and loss function. Results indicated that GFCC performed better among the other cepstral coefficients. Hyperparameter tuning with the Adamax optimizer resulted in a median accuracy of 68.1%. Among loss functions, Binary Cross-entropy (BCE) showed promising results with a median accuracy of 68% and minimal accuracy variation of 9%. Despite all the batch sizes providing comparable accuracy, batch size 32 showed a lower variation of 4.5%. Thus, the EfficientNet model is trained using the Adamax optimizer with a learning rate of 0.001 using BCE loss in batch sizes of 32. Five-fold cross-validation of this model gives 73% mean accuracy and precision rate of 0.73. As GFCC is capable of capturing transient sounds which are crucial for finding distinguishing patterns in AD, the EfficientNet architecture with GFCC gives a better performance in differentiating Alzheimer’s condition. Thus, this framework could be used for acoustic signal-based classification of Alzheimer’s disease.