An improved ViT model for music genre classification based on mel spectrogram.

Area of Science:

• Artificial Intelligence
• Music Information Retrieval
• Machine Learning

Background:

• Automated music genre classification is crucial for enhancing user experiences and managing music libraries.
• Existing methods may not fully capture complex features within music audio signals.

Purpose of the Study:

• To propose an improved Vision Transformer (ViT) model for more accurate music genre classification.
• To enhance feature extraction from Mel spectrograms by combining Convolutional Neural Networks (CNNs) and Transformers.
• To improve classification precision using a channel attention mechanism.

Main Methods:

• Utilized an improved Vision Transformer (ViT) architecture.
• Integrated Convolutional Neural Networks (CNNs) with Transformers for feature extraction.
• Incorporated a channel attention mechanism to amplify inter-channel differences in Mel spectrograms.
• Evaluated the model on the GTZAN dataset.

Main Results:

• The proposed model achieved an accuracy of 86.8% on the GTZAN dataset.
• Demonstrated superior performance in extracting comprehensive music genre features compared to previous approaches.
• The channel attention mechanism contributed to more precise classification.

Conclusions:

• The improved ViT model offers a more accurate and efficient method for music genre classification.
• This approach enhances the ability to understand and categorize diverse music genres.
• The findings pave the way for advancements in music information retrieval systems.