Unsupervised Evaluation of Deep Audio Embeddings for Music Structure Analysis

arXiv cs.AI · 6 hours ago

Researchers evaluated nine pre-trained deep audio models on music structure analysis using unsupervised segmentation methods without labeled training data. The Correlation Block-Matching algorithm proved most effective among three tested segmentation approaches, and modern deep embeddings outperformed traditional spectrogram baselines inconsistently. The study proposes stricter evaluation standards for music structure analysis by adopting trimming methods to prevent artificial inflation of performance metrics.