在數據驅動的時代，邊緣運算和人工智慧的發展正重新塑造各行各業的運作模式。隨著物聯網設備的普及和數據生成量的激增，傳統的集中式雲計算架構面臨著巨大的挑戰，如高延遲、頻寬限制和數據安全問題。因此，將計算和數據處理推向網路邊緣，以實現即時響應和高效處理，成為了一種重要的技術趨勢。 然而，邊緣運算中的挑戰在於資源限制——邊緣設備通常擁有較低的計算能力和儲存空間。這就需要精簡化的AI模型來在這些有限的設備上高效運行。AI模型的精簡化又不失精確度，可以有效地減少模型的計算需求和存儲佔用，從而使得即使在邊緣設備上也能實現高效的數據處理和即時響應。本文探討了在關鍵字檢測中如何實現AI模型的輕量化，同時保證其精確度，以適應邊緣設備的需求。 ;In the data-driven era, the development of edge computing and artificial intelligence is reshaping the operational models across various industries. With the proliferation of IoT devices and the surge in data generation, traditional centralized cloud computing architectures are facing significant challenges, such as high latency, bandwidth limitations, and data security concerns. As a result, shifting computation and data processing toward the network edge to achieve real-time responsiveness and efficient processing has become an important technological trend. However, edge computing presents challenges due to resource constraints—edge devices typically possess limited computing power and storage capacity. This necessitates the use of streamlined AI models that can run efficiently on these constrained devices. The simplification of AI models, without compromising accuracy, can effectively reduce computational demands and storage usage, thereby enabling efficient data processing and real-time response even on edge devices. This paper explores how to achieve model lightweighting for keyword spotting while maintaining accuracy, in order to meet the needs of edge devices.
