A music structure analysis method based on beat feature and improved residual networks.

Area of Science:

• Music Information Retrieval
• Computational Musicology
• Machine Learning for Audio Analysis

Background:

• Traditional music structure analysis (MSA) methods struggle with inadequate audio feature representation and limited model generalization.
• Accurate music segmentation and labeling are crucial for understanding musical pieces.

Purpose of the Study:

• To develop an advanced MSA method addressing limitations in current approaches.
• To improve the accuracy of boundary detection and segment labeling in music.

Main Methods:

• Refactoring music structure labels into 9 types, refined to the beat level.
• Employing beat-wise feature extraction incorporating diverse acoustic features for precise segmentation.
• Utilizing a Resnet-34 with a self-attentive mechanism for beat-wise category prediction.
• Implementing a post-processing step for filtering predicted labels.

Main Results:

• The proposed method achieves state-of-the-art performance on the SALAMI-IA dataset.
• Demonstrates a 3 percentage point improvement over the previous optimal method on HR3F.
• Outperforms existing advanced methods on PWF and Sf metrics.

Conclusions:

• The beat feature fusion and improved residual network effectively enhance music structure analysis.
• The method offers superior accuracy in both boundary detection and segment labeling tasks.
• This approach represents a significant advancement in computational musicology and music information retrieval.