발간년도 : [2023]
| 논문정보 |
|
| 논문명(한글) |
[Vol.18, No.6] Analysis of Machine Learning-based Environmental Sound Classification Using Explainable Artificial Intelligence(XAI) |
|
| 논문투고자 |
Wongeun Oh |
|
| 논문내용 |
Sound classification using machine learning is a technology that can be effectively applied to variousfields. Several methods, such as adapting more complex models or improving pre-processing stages, havebeen attempted to increase the recognition performance of sound classifiers, and as a result, algorithmswith high recognition rates are continuously being developed. However, it is often impossible to clearlyexplain why the performance has improved, and this is due to the black box characteristics of machinelearning models. In this paper, we applied an explainable artificial intelligence(XAI) technique to themachine learning sound classifier and analyzed which part of the spectrogram of the sound to beclassified affects the results. The XAI is applied in various applications, but only a few studies areconducted for sound recognition compared to image data analysis. The XAI technique used in this paperis LIME, a surrogate model algorithm that can be applied regardless of the model or learning method.The dataset used in the experiment is UrbanSound8K, which consists of 10 types of urban environmentalsounds, and the machine learning model to handle it consists of a VGG16-based transfer learning model.The experiment uses the LIME algorithm to derive the spectrogram area most significantly affecting theclassification results. Through this, we tried to find out the recognition patterns of the spectrogram whensound was correctly recognized and when it was not. The analysis results were different depending onthe sound. Complex sounds and impulsive sounds showed some common classification patterns, butsounds mixed with several sounds did not show a clear pattern. The results of the LIME analysisdiffered depending on the type of sound, and these could be improved by adjusting the LIME parametersor applying other XAI techniques. In the future, we plan to refine the taxonomy of the sounds furtherand analyze a wider variety of sounds using different audio datasets. |
|
| 첨부논문 |
|
|
|
|
|
자료실 > 논문지 
