Search Results for « Lightweight Speaker Verification »

This paper proposes a lightweight architecture for speaker recognition and verification that integrates Reservoir Computing (RC) with multi-scale and multi-head attention mechanisms. The fixed recurrent reservoir captures long-term temporal dynamics, while the attention modules selectively aggregate speaker-discriminative information across multiple temporal resolutions. By restricting learning to compact projection and attention layers, the proposed RC-Attention framework considerably reduces the number of trainable parameters while maintaining strong discriminative capability. The proposed architecture is evaluated on three widely used benchmarks, VoxCeleb1, LibriSpeech, and CNCeleb1, covering both in-domain and cross-corpus open-set verification scenarios. Under the standard VoxCeleb2-dev/VoxCeleb1-O evaluation protocol, RC-Attention achieves an Equal Error Rate (EER) of $2.71\%\pm0.05\%$ with approximately 2.1M parameters. An extensive ablation study demonstrates that the reservoir is an essential component of the proposed architecture, to achieve the intended efficiency-performance balance, and that multi-scale temporal aggregation provides the largest contribution to verification performance. Replacing the fixed reservoir with trainable recurrent modules (GRU or LSTM) further improves EER, but requires substantially more trainable parameters, more than twice the training time, nearly twice the computational cost, and significantly higher inference latency. Cross-corpus experiments further demonstrate that the proposed architecture generalizes well under domain and language mismatch, consistently improving performance on LibriSpeech and CNCeleb1 under a restricted-data training regime. Overall, the proposed RC-Attention framework offers an attractive trade-off between verification performance, model complexity, and computational efficiency, making it a promising solution for resource-constrained and real-time speaker verification applications.