Please use this identifier to cite or link to this item: https://hdl.handle.net/10356/158686
Title: A lightweight convolutional neural network for KWS
Authors: Wang, Yumengmeng
Keywords: Engineering::Electrical and electronic engineering
Issue Date: 2022
Publisher: Nanyang Technological University
Source: Wang, Y. (2022). A lightweight convolutional neural network for KWS. Master's thesis, Nanyang Technological University, Singapore. https://hdl.handle.net/10356/158686
Abstract: Keyword spotting has been widely used in smart homes and mobile devices, where the goal is to achieve high accuracy with low latency and small computational effort. Convolutional neural networks are not only performed well in extracting speech features, but also show good robustness in the face of noise interference. This dissertation proposes a convolutional neural network using Google's speech command as the dataset, with a trade-off between high accuracy and small memory. Existing techniques were applied to modify the network to obtain the results of the best performing model. Further, the training set is noise-loaded to show the effect of noise on the audio files, and the performance of the KWS-Net under three levels of noise is demonstrated. Finally, some advanced deep learning networks are used for speech recognition experiments on the same dataset, including ResNet, DenseNet, MobileNet. The results section compares the test results of the proposed network with these advanced networks in speech recognition, highlighting the good performance achieved by the KWS-Net in terms of high accuracy and low number of parameters.
URI: https://hdl.handle.net/10356/158686
Fulltext Permission: restricted
Fulltext Availability: With Fulltext
Appears in Collections:EEE Theses

Files in This Item:
File Description SizeFormat 
Dissertation_WangYuMengmeng_Final.pdf
  Restricted Access
3.92 MBAdobe PDFView/Open

Page view(s)

58
Updated on Dec 5, 2022

Download(s)

3
Updated on Dec 5, 2022

Google ScholarTM

Check

Items in DR-NTU are protected by copyright, with all rights reserved, unless otherwise indicated.