Hardware-friendly neural network design and optimization for low power IOT applications

As speech becomes a popular way for human to interact with electronic devices in recent years, it leads to interests to apply machine learning in speech related applications, such as sound classification, speech recognition and so on. One of the exciting applications is to develop keyword spottin...

Full description

Saved in:
Bibliographic Details
Main Author: Wang, Yingfeng
Other Authors: Goh Wang Ling
Format: Final Year Project
Language:English
Published: Nanyang Technological University 2022
Subjects:
Online Access:https://hdl.handle.net/10356/157996
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Nanyang Technological University
Language: English
Description
Summary:As speech becomes a popular way for human to interact with electronic devices in recent years, it leads to interests to apply machine learning in speech related applications, such as sound classification, speech recognition and so on. One of the exciting applications is to develop keyword spotting (KWS) module using neural network. The KWS module is acting as a switch to activate a downstream system, for example, a speech recognition system after certain keywords have been detected. In actual application, a high accuracy KWS which is able to identify the keyword with the existence of background noises is desired for a smooth user experience. Thus, this project aims to design a hardware-friendly and noise-robust neural network for KWS, expecting to classify 10 keywords along with “silence” and “unknown” class. A final LSTM model with 4-bit quantization and k=9 pruning shows 91.74% accuracy on clean audio, with a model size of 7.9KB. Compared to other state-of- the-art KWS architectures classifying the same number of keywords, this work is able to achieve a 2-4% higher accuracy for both clean and noisy audios, as well as a size reduction of at least 29%.