Reservoir Computing with Multi-Scale Attention for Efficient Speaker Verification

This paper proposes a lightweight architecture for speaker recognition and verification that combines Reservoir Computing (RC) with multi-scale and multi-head attention. A fixed recurrent reservoir models long-term temporal dynamics, while attention selectively aggregates speaker-discriminative information across temporal resolutions, restricting learning to compact layers and reducing the number of trainable parameters. Evaluations on VoxCeleb1, LibriSpeech, and CNCeleb1, including cross-corpus open-set scenarios, show competitive Equal Error Rate (EER) compared to state-of-the-art models with at least 3.5x fewer parameters. Ablation results indicate that reservoir dimensionality and multi-scale aggregation drive performance, while attention improves robustness and feature integration. Strong generalization under domain mismatch further highlights the effectiveness of the proposed approach, making it suitable for resource-constrained and real-time applications.