Voice detection with spiking convolutional neural network for smart sensor applications

• Main Author: Leow, Cong Sheng
• Other Authors: Goh Wang Ling
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Electrical and electronic engineering::Computer hardware, software and systems; Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/153415
• Institution: Nanyang Technological University

Audio detection on the edge can bring great value in various areas, be it at home, in healthcare sectors, or even in the industry. Smart sensors, therefore, play an important role in enabling that, and these sensors require high intelligence and low power consumption. While conventional deep learning approaches have evolved tremendously and have reached exceptional performance in tasks such as audio detection, it is challenging to implement highly complex neural networks without requiring high computational resources. Neuromorphic computing is an emerging field of study which seeks to achieve the efficiency and performance of the biological brain through the incorporation of biological-plausible mechanisms and emulation into electronic computing systems. Spiking neural network (SNN) is the next-generation neural network used in many of today’s neuromorphic systems. By modelling neurons and learning mechanisms closely to how the biological brain operates, SNN seeks to achieve greater efficiency than a conventional neural network. In this project, a spiking convolutional neural network (SCNN) was implemented by using a spiking layer consisting of leaky-integrate-and-fire (LIF) neurons with a convolutional neural network. A study of the audio processing techniques and the neuron parameters in the SCNN was done to achieve optimal performance when compared with a deep learning approach. The SCNN achieved an accuracy of over 80% while using fewer layers than a high-performance deep convolutional neural network. The proposed model provides a better understanding of SNN for audio detection and paves the way for hardware implementation which could be efficient and effective for smart sensor applications.