<p>Automatic speaker identification aims to recognize individual speakers from a mixture of voices and background sounds. This task is challenging due to the high variability of speech signals, emotional fluctuations that alter vocal characteristics, and the limitations of two-dimensional convolutional neural networks (2D-CNNs) in capturing both spatial and temporal dependencies. This paper proposes a data preprocessing method that enhances the performance of speaker identification systems by preserving speaker-specific information while efficiently handling audio signals of varying lengths. In addition, a hybrid architecture is introduced for temporal spectrograms, combining three-dimensional convolutional layers (Conv3D) with temporally distributed two-dimensional convolutional layers (Conv2D). This design allows the model to capture spatial and temporal dynamics simultaneously, improving its ability to differentiate between speakers. Experimental results confirm the effectiveness of the proposed approach. The model achieves significant improvements in accuracy and robustness under conditions of acoustic mismatch, accent, and emotional variation. Overall, the method is not sensitive to the speaker’s emotional state, accent, or language. The proposed system reaches an accuracy of up to 99% on the Aishell-1 database and more than 98% on emotional datasets such as RAVDESS.</p>