Smart computers : can you recognize these sounds?
Support vector machines (SVMs) is a common form of sound classification. This paper aims to employ SVM methods to classify 13 kinds of sound events in a noisy environment. This paper finds that a more than 80% of classification accuracy is achievable in the controlled environment. This is supported...
Saved in:
Main Author: | |
---|---|
Other Authors: | |
Format: | Final Year Project |
Language: | English |
Published: |
2015
|
Subjects: | |
Online Access: | http://hdl.handle.net/10356/63579 |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Institution: | Nanyang Technological University |
Language: | English |
Summary: | Support vector machines (SVMs) is a common form of sound classification. This paper aims to employ SVM methods to classify 13 kinds of sound events in a noisy environment. This paper finds that a more than 80% of classification accuracy is achievable in the controlled environment. This is supported by real-time testing indications. This paper starts with an introduction to SVMs and how it classifies sound events by extracting features, mapping them onto a feature space illustrated in 2-D form. As SVMs are linear two-class classifiers, this method has to be expanded to a multi-class SVM via the introduction of kernels; Gaussian kernel in particular. Kernel parameters affect how the optimum boundaries are drawn and how the hyperplanes are constructed. The data collection process involving 50 personnel doing 13 classes of sound events is detailed together with the equipment used and the procedure. Data processing via MATLAB, like model training, testing and real-time testing is explained. 3 models were trained and tested, each having different numbers of audio files in training and testing. The impact on the accuracy is examined in the form of a confusion matrix. A discussion on the real time testing is done at the end, with limitations of the models explained and future work suggested. |
---|