| Title |
Mixed-type data augmentations for environmental sound classification / |
| Authors |
Turskis, Tadas ; Teleiša, Marius ; Buckiūnaitė, Rūta ; Čalnerytė, Dalia |
| Full Text |
|
| Is Part of |
CEUR workshop proceedings: IVUS 2023: Information society and university studies 2023: proceedings of the 28th international conference on information society and university studies (IVUS 2023) Kaunas, Lithuania, May 12, 2023 / edited by: A. Lopata, T. Krilavičius, I. Veitaitė, A. García-Holgado.. Aachen : CEUR-WS. 2023, vol. 3575, p. 184-194.. ISSN 1613-0073 |
| Keywords [eng] |
environmental sound classification ; mixed augmentation ; Mel-frequency cepstral coefficients |
| Abstract [eng] |
The goal of environmental sound classification is to accurately identify and classify sounds in order to provide valuable insights about the environment. The classification task can be solved by training machine learning models, such as convolutional neural networks, on a dataset of labeled sound samples. Due to the small size of available datasets in this field, time-consuming and expensive labeling process, data augmentations have become a popular practice to artificially generate additional data. The purpose of this study is to analyze whether using Mixed-Type data augmentations improves the classification performance compared to results with no augmentations. Mixed-Type data augmentation methods were evaluated on ESC-50 and UrbanSound8K datasets for the pretrained ResNet-18 model with extracted mel-frequency cepstral coefficients as feature inputs. Results for both datasets show that data augmentations can improve model performance with certain mixup probabilities and coefficients but specific methods and parameters used may vary for each dataset and task. |
| Published |
Aachen : CEUR-WS |
| Type |
Conference paper |
| Language |
English |
| Publication date |
2023 |
| CC license |
|
