Design of a DNN-based operator on edge device for keyword spotting

• Published in: Advances in computational intelligence Vol. 5; no. 2; p. 2
• Main Authors: Wei, Chan Kok, Nugroho, Hermawan
• Format: Journal Article
• Language: English
• Published: Cham Springer International Publishing 01.06.2025; Springer Nature B.V
• Subjects: Accuracy; Algorithms; Artificial Intelligence; Artificial neural networks; Computational Intelligence; Constraints; Critical components; Datasets; Deep learning; Edge computing; Efficiency; Embedded systems; Engineering; Ensemble learning; Floating point arithmetic; Inference; Internet of Things; Keywords; Machine Learning; Neural networks; Noise; Original Article; Parameters; Real time; Test sets; User experience; Voice recognition; MobileNetV2; Convolutional neural networks (CNN); Ensemble learning; Knowledge distillation; Keyword spotting; Edge computing
• ISSN: 2730-7794; 2730-7808
• DOI: 10.1007/s43674-025-00080-2

Keyword spotting (KWS) is a critical component of voice-driven smart-device applications, requiring high accuracy, sensitivity, and responsiveness to deliver optimal user experiences. Given the always-on nature of KWS systems, minimizing computational complexity and power consumption is essential, particularly for battery-powered edge devices with constrained resources. In this paper, we propose a compact and highly efficient convolutional neural network (CNN) for edge-based KWS tasks, using the Google Speech Commands (GSC) V2 dataset for training and evaluation. Our model employs modified MobileNetV2 architecture, optimized via knowledge distillation from an ensemble of high-performing CNN models. Experimental results demonstrate that the proposed model achieves 94.48% accuracy on clean test data and significantly outperforms existing state-of-the-art edge models on challenging noisy test sets, reaching 86.38% accuracy. The proposed CNN maintains this superior performance with only 73.8K parameters and 19.5M floating-point operations (FLOPs)—approximately three times fewer FLOPs and substantially fewer parameters than previously reported edge-focused KWS models. Moreover, when evaluated on a realistic and challenging external Kaggle test set, the proposed model shows excellent generalization with 88.38% accuracy, surpassing baseline depthwise separable CNN (DS-CNN) approaches. Upon practical deployment on a widely used embedded computing platform, our optimized model achieved fast inference times between 11 ms and 14 ms per sample, outperforming existing baseline methods and confirming its suitability for real-time applications. This study highlights the successful integration of model compression techniques, including ensemble learning and knowledge distillation, to achieve breakthrough performance improvements in accuracy, robustness to noise, computational efficiency, and inference speed, thereby advancing the practical deployment of high-performance KWS solutions on resource-constrained edge devices.