Research on music audio chord feature extraction and similarity calculation based on transformer and deep metric learning

Aiming at the problem that existing methods have difficulty balancing chord recognition accuracy and feature discrimination, this paper proposes the Transformer Chord Metric Extractor (TCME). The model effectively captures the long-range harmony dependence of music by designing a local–global hierarchical attention mechanism. The experimental results show that the weighted accuracy of the TCME model in the chord recognition task reaches 0.90, which is better than 0.86 for the standard Transformer; in the music similarity retrieval task, the average Top-5 accuracy is as high as 0.89, a significant improvement over the Weaviate baseline.