Self-Supervised Transfer Learning from Natural Images for Sound Classification

We propose the implementation of transfer learning from natural images to audio-based using self-supervised schemes. Through learning, convolutional neural networks (CNNs) can learn general representation without labels. In this study, a network was pre-trained with (ImageNet) via learning; subsequently, it fine-tuned on target audio samples. Pre-training scheme significantly improved sound classification performance when validated following benchmarks: ESC-50, UrbanSound8k, and GTZAN. The achieved similar level accuracy as those supervised method that require Therefore, we demonstrated contributes improvements in audio-related tasks, is adequate for pre-training terms simplicity effectiveness.