Energy efficient voice detection with spiking neural network for smart sensor applications

• 主要作者: Tan, Chiah Ying
• 其他作者: Goh Wang Ling
• 格式: Final Year Project
• 語言: English
• 出版: Nanyang Technological University 2020
• 主題: Engineering::Electrical and electronic engineering::Applications of electronics
• 在線閱讀: https://hdl.handle.net/10356/140766

This project aims to design an algorithm for classifying voice commands for smart sensor devices, allowing them to be deployed in environments with poor network connectivity. The goal of the algorithm is to be energy efficient, so that the neural network that does the voice command classification can be executed locally on the device, without the need for cloud computing. This would be done by utilising a new generation of neural networks – spiking neural networks. The spiking neuron’s unique characteristic of transmitting information with the use of electrical spikes will be used to convert the voice samples into a sparse form, whereby the spikes represent segments of the voice sample with critical information, while the rest of the sample will be converted to zeros. This allows the hardware device to conserve a significant amount of energy, since it is able to maintain a state of rest in tandem during the rest periods of the spiking neuron. The spiking neural network was created with the Leaky integrate and fire model, and two classification tasks were undertaken. One was to classify voices based on gender, while the other was to classify the voice commands based on the command issued. The algorithm was trained using the Google Commands dataset, achieving 91% and 98% on the gender and command word classification task, respectively.