Development and Optimization of an Ultra-lightweight Deep Spoken Keyword Spotting Model for FPGA Acceleration

Automatic speech recognition (ASR) has become one of the most advanced and studied domains in human-facing machine learning applications. Spoken Keyword Spotting (KWS), a subset of ASR, is a technology that enables systems to detect specific keywords or phrases in spoken language. Modern machine learning models, such as deep neural networks, have significantly advanced the performance and accuracy of KWS systems. However, they often demand substantial computational resources and introduce latencies that limit their real-time applicability and offline use. This has become a tremendous problem where faster and more efficient processing methods dominate and better meet industry demands. To address this challenge, this paper developed a lightweight 1-Dimensional convolutional neural network based on the Mel-frequency cepstral coefficient input and compressed it with quantization and pruning for deployment onto FPGA hardware. The developed model achieved near state-of-the-art performance with far fewer parameters and a simpler architecture than comparable models in literature, and it showed significant model compression with only minor accuracy degradation. This paper also leveraged FPGAs as the hardware deployment strategy to evaluate their effectiveness as inference accelerators for KWS models based on their resource utilization and latency performance improvements.