@article{CRSIJ26000357,
                  author = {Oni, Oluwabunmi Ayankemi, Ayeni, Joshua Ayobami, Makinde, Oladayo Ezekiel, Ojo, Olufemi Samuel},
                  title = {Response-Based Knowledge Distillation for Efficient Multilingual Keyword Spotting: A Distilled Depthwise Separable CNN Approach},
                  journal = {Cosmo Research and Science International Journal},
                  year = {2026},
                  volume = {2},
                  number = {2},
                  pages = {134-141},
                  issn = {3108-1584},
                  url = {https://cosmorsij.com/published/CRSIJ26000357.pdf},
                  abstract = {Keyword spotting (KWS) on resource-constrained devices remains challenging, especially for low-resource languages with limited labelled data. This paper proposes a response-based knowledge distillation framework that transfers discriminative knowledge from a high-capacity Res Net teacher to a lightweight Depth wise Separable Convolutional Neural Network (DSCNN) student. Using Mel-Frequency Cepstral Coefficients and a hybrid loss combining cross-entropy and temperature-scaled Kullback–Leibler divergence, the distilled model (D-DSCNN) achieves 96.40% test accuracy while retaining the same memory footprint (≈132 KB) and computational cost (1,728 MACs) as the baseline DSCNN. Experiments on a multilingual crime-related keyword dataset demonstrate that knowledge distillation significantly improves generalization without increasing model complexity, making the approach suitable for always-on edge deployment.},
                  keywords = {Keyword spotting, Knowledge distillation, DSCNN, Low-resource languages, Edge devices, Multilingual},
                  month = {September}
        }