Speaker recognition is essential in applications such as security systems, forensic analysis, and voice-activated technology. With advances in deep learning, speaker identification has become increasingly accurate and efficient. This study presents a methodology for identifying speakers based on vocal features, using the publicly available Speaker Recognition Dataset comprising 18,544 audio samples from 38 distinct speakers. We extracted mel-frequency cepstral coefficients (MFCCs) as feature vectors, implementing both 20 and 40 MFCC configurations to capture key vocal characteristics. These features were used to train three different deep learning models, convolutional neural networks (CNNs), long short-term memory (LSTM) networks, and recurrent neural networks (RNNs) to evaluate model performance in classifying speakers. Using tenfold cross-validation, the models were trained and tested with the Speaker Recognition Dataset, allowing for unbiased performance comparisons. In the 20 MFCC configuration, the CNN model achieved the highest test accuracy 94.28%, surpassing the LSTM 94.10% and RNN 94.22% models. Similarly, in the 40 MFCC configuration, the CNN model showed the best performance with a test accuracy of 93.93%, outperforming both the LSTM model at 93.11% and the RNN model at 91.54%. These findings underscore the CNN model’s effectiveness for speaker recognition tasks on this dataset, particularly when using MFCCs, though further research with various datasets is recommended to assess the broader applicability of these models.