Scholarlink

Current state-of-the-art (SOTA) music genre classification models exhibit significant Western bias due to predominant training on Western datasets, causing suboptimal performance on non-Western genres with distinct characteristics. This research introduces a contrastive pretraining approach targeting Greek music genres, including Laiko, Rebetiko, and Entehno. Three deep learning architectures are tested: custom Convolutional Neural Networks (CNNs), hybrid Convolutional Recurrent Neural Networks (CRNNs), and pretrained transformers. Models first undergo contrastive pretraining on 13,000 unlabeled Greek tracks before fine-tuning on 1,033 labeled tracks. Macro-F1 (unweighted mean of per-class F1 scores) for the custom CNN and CRNN increased by 0.04 and 0.01, respectively, after the contrastive step and reduced convergence time, demonstrating its effectiveness in data-scarce scenarios. However, no measurable improvement was observed on the transformers, likely due to restricted batch size for effective contrastive learning.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Mitigating Western Bias in Music Genre Classification: A Contrastive Learning Approach for Greek Music

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • George Kritsovas,
  • Lorenzo Garbagna,
  • Lakshmi Babu Saheer,
  • Mahdi Maktabdar Oghaz

摘要

Current state-of-the-art (SOTA) music genre classification models exhibit significant Western bias due to predominant training on Western datasets, causing suboptimal performance on non-Western genres with distinct characteristics. This research introduces a contrastive pretraining approach targeting Greek music genres, including Laiko, Rebetiko, and Entehno. Three deep learning architectures are tested: custom Convolutional Neural Networks (CNNs), hybrid Convolutional Recurrent Neural Networks (CRNNs), and pretrained transformers. Models first undergo contrastive pretraining on 13,000 unlabeled Greek tracks before fine-tuning on 1,033 labeled tracks. Macro-F1 (unweighted mean of per-class F1 scores) for the custom CNN and CRNN increased by 0.04 and 0.01, respectively, after the contrastive step and reduced convergence time, demonstrating its effectiveness in data-scarce scenarios. However, no measurable improvement was observed on the transformers, likely due to restricted batch size for effective contrastive learning.