Keyword Spotting for Edge AI: Lightweight Decoding and Training-Free Debiasing

We propose a lightweight keyword spotting pipeline tailored for real-time edge AI applications. It addresses the latency bottleneck of auto-regressive decoders in image captioning by replacing the sentence-generation stage with a single-layer keyword spotting head, while reusing a strong pretrained encoder. Furthermore, we introduce a training-free debiasing method-z-ranking-that standardizes per-class scores to correct frequency bias with negligible runtime cost. On the MS COCO 2014 validation set, our z-ranking method improves micro-F1 from 0.2535 to 0.2885 (a 13.8% relative gain) without retraining. The end-to-end latency is reduced by a factor of $1.92 \times$ compared to the full captioning path, and the decoder itself is accelerated by over $\mathbf{1 0 0} \boldsymbol{\times} \boldsymbol{(} \mathbf{0. 1 0 5 ~ m s}$ vs. $\boldsymbol{\approx} \mathbf{1 1. 5} \mathrm{ms})$. These results demonstrate that the proposed pipeline is a practical and lightweight solution suitable for real-time edge AI applications.