School of Engineering News

Revolutionizing Edge AI: Award-Winning FPGA-Optimized Deep Learning Model Advances Keyword Spotting Technology

November 27, 2024

Revolutionizing Edge AI: Award-Winning FPGA-Optimized Deep Learning Model Advances Keyword Spotting Technology

GVSU student and faculty duo Trysten Dembeck and Dr. Chirag Parikh are making waves after their paper titled "Development and Optimization of an Ultra-Lightweight Deep Spoken Keyword Spotting Model for FPGA Acceleration" won the Best Paper Finalist award at the CAINE 2024 Conference. The paper highlighted significant advancements in Spoken Keyword Spotting (KWS) technology, a critical component of speech-to-text systems.

The research focuses on making high-performance machine learning models practical for use on compact, embedded devices. It strikes a balance between accuracy, size, and speed to create a lightweight yet highly effective keyword spotting (KWS) model based on a 1-D Convolutional Neural Network. By using model compression and advanced training methods, the model was successfully adapted for FPGA hardware, overcoming the memory and speed limitations of edge devices. The ability of FPGAs to process data quickly and in parallel allowed for real-time KWS performance with incredibly fast response times. This achievement combines high accuracy, quick performance, and a compact model design that has met or exceeded the performance of similar systems in recent literature.

This research represents a pivotal step forward for the future of KWS technology, demonstrating how cutting-edge and increasingly complex machine learning models can be optimized for deployment onto hardware platforms. Many smart home devices that use this technology require an internet connection to have access to highly accurate speech-to-text models. This work takes strides towards removing the requirement of an internet connection to have accurate and fast responding smart devices in more applications than previously possible. By showcasing an FPGA design capable of being implemented into an Application-Specific Integrated Circuit (ASIC), the work highlights the potential for creating compact, energy-efficient devices without sacrificing performance. Additionally, as microcontroller manufacturers continue to integrate hardware acceleration ICs into their designs, the possibilities for deploying high-performance KWS solutions at the edge become increasingly feasible. These advancements are set to drive faster, more reliable speech-based interfaces, solidifying their role in next-generation technologies and hands-free applications.

Share this news story

View More School of Engineering News



Page last modified November 27, 2024