The purpose of this project is to investigate how urban audio samples in computer-readable forms might be classified into the appropriate classes using an image transfer learning model. As a baseline, the EfficientNet-B1 model is used, along with Log-mel Spectrogram feature representations. To improve the baseline results,various audio data augmentation techniques are investigated. The two notebooks available are for the ESC-50/ESC-10 dataset and the UrbanSound8k dataset.
- 94.25% on ESC-10
- 85.9% on ESC-50
- 76.5% on UrbanSound8k
- The model architecture was adapted from J.Kim in the dcase 2020 challenge
If you have any feedback, please reach out to me at katrinemadfathy@gmail.com
-
Experiment with larger network
-
Experiment with class conditional data augmentation