Noise-robust speaker recognition using an enhanced multi-layer CNN with attention and recurrent modeling

Abstract

Automated speaker recognition systems often degrade in noisy, emotional, and cross-domain conditions due to the limited robustness of traditional statistical and shallow learning models. While CNN-based approaches enhance local time frequency representations, many still lack effective mechanisms for modeling long-term temporal dependencies and contextual salience required for reliable real-world deployment. To address these challenges, this paper proposes an Enhanced Multi-Layer Convolutional Neural Network (EML-CNN) for robust speaker identification. The framework integrates dense convolutional blocks, Gated Recurrent Units (GRUs), and an attention mechanism to learn highly discriminative speaker representations. The proposed model is evaluated on three benchmark datasets: IEMOCAP (emotional speech), THUYG-20 (dialectal diversity), and VoxCeleb (noisy multi-speaker conditions), achieving accuracies of 99.4%, 96.1%, and 96.4%, respectively. Robustness is further enhanced through advanced audio augmentation, including additive Gaussian noise (SNR 10–30 dB), temporal shifting, frequency stretching, and pitch variation. Experimental results demonstrate that EML-CNN consistently outperforms traditional and recent state-of-the-art methods, confirming its suitability for real-world applications such as secure access control, forensic speaker verification, and human-computer interaction.