| Abstract:Speech Emotion Recognition (SER) is increasingly deployed in edge scenarios, where real-time performance and energy efficiency are critical. However, the quadratic computational complexity of Transformers limits their applicability, while existing lightweight methods with fixed temporal shift patterns fail to capture dynamic emotional variations. To address these challenges, we propose Edge-ADT, an adaptive shift-based model tailored for edge deployment. Edge-ADT replaces multi-head attention with parameter-efficient shift operations and introduces a lightweight decision network to dynamically generate channel-wise shift coefficients, enabling adaptive temporal feature alignment with linear complexity.Experimental results demonstrate that, compared with standard Transformers, Edge-ADT reduces parameters by 27.31% and FLOPs by 27.52%. Compared with temporal shift methods, it improves Unweighted Accuracy (UA) by 0.88% and Weighted Accuracy (WA) by 0.87%. On the NVIDIA Jetson Nano platform, Edge-ADT achieves an inference latency of 300.3 ms per sample with a power consumption of 6.1 W, delivering a 66.5% speedup over the baseline. Under 0 dB white noise, the UA degrades by only 10.53%, outperforming the 14.28% drop of Transformers, demonstrating superior robustness to noise.These results validate the effectiveness and efficiency of the proposed adaptive shift mechanism for real-world edge SER applications. |
点此下载全文
位访问者,今日访问
