Speaker identification is determining a speaker’s identity from an audio sample. Recent advances in deep learning have enabled significant improvements over traditional methods. In this work, we develop a speaker identification system utilizing multiple neural architectures, including Long Short-Term Memory (LSTM) networks, convolutional-recurrent hybrids (CNN-LSTM), transformers, an LSTM with attention, and a novel feature fusion model that combines CNN and LSTM branches. We train and evaluate these models on a dataset of 20 speakers with augmented noisy data. Short-time Fourier transform and Mel-frequency cepstral coefficient (MFCC) features are extracted to represent audio signals. The deep models are trained on these features and compared in terms of accuracy, precision, recall, and F1-score. Our experiments demonstrate that hybrid architectures, such as CNN-LSTM and transformer-based models, significantly outperform a baseline LSTM (87%) with a high identification accuracy of 97%. We discuss the impact of noise augmentation and attention mechanisms on performance. Overall, the results demonstrate the effectiveness of deep learning for speaker identification, providing insights into the benefits and limitations of hybrid neural architectures. Future work will explore larger datasets, more advanced fusion strategies, and transformer-based embedding models to improve speaker identification performance further.