발간년도 : [2021]
| 논문정보 |
|
| 논문명(한글) |
[Vol.16, No.6] Effect of Music on the Performance Degradation of Sound Classification Using Deep Learning |
|
| 논문투고자 |
Wongeun Oh |
|
| 논문내용 |
Sound classification is a computer technology that involves learning to classify sounds and to predict the category of that sound. Recently, the machine learning based approach is being actively conducted for improving recognition accuracy. In this approach, a deep neural network is trained using a sound dataset, and then the actual sound is applied to identify the sound category. In the identification stage, the recognition accuracy of the machine learning is degraded due to the ambient noise. In other words, unlike the experimental environment, various sounds are input into the microphone along with the target sound. Since these ambient sounds are not trained, they could lower the classification performance. However, there are only a few research results on the relation between the noise and the recognition performance despite of the practical importance. In this paper, we study the performance degradation of sound classification in the space where the music is playing with considering the music genre and the volume level. For this, we use CNN, UrbanSound8K dataset consisting of 10 kinds of environmental sounds, and GTZAN data set containing 10 kinds of music genres. First, CNN is trained to recognize the sounds of UrbanSound8K, and then five songs for each genre were selected from GTZAN and mixed to the UrbanSound8K so that the signal-to-noise ratio are -20dB, 0dB, 5dB, 10dB, and 20dB. Then we test the accuracy with the mixed sound input and compare with the noise-free target sound. As a result, there is 2.8% to 22% difference in the recognition accuracy by music genre and sound level. The result show that the SNR should be 20dB or more in order for music not to have a significant effect on the recognition accuracy. |
|
| 첨부논문 |
|
|
|
|
|
자료실 > 논문지 
