用条件相似度网络对 CLAP、MuQ、MERT 冻结嵌入做事后解耦
Exploring Post-Hoc Disentanglement of Audio Foundation Models for Music Similarity
CLAP、MuQ 这些音频模型的嵌入是纠缠在一起的，这篇论文证明加一层很小的可训练门控就能把属性拆开，检索效果还顺带提升。做音乐检索或者想精细控制属性的工程师可以看看实现细节，不用重训骨干，成本很低。
arXiv 论文（编号 2611.14415v1）提出对音频基础模型做事后解耦（post-hoc disentanglement），将条件相似度网络（CSN）适配到冻结的预训练嵌入上，替代从零训练骨干模型的做法。实验覆盖 CLAP、MuQ、MERT 三个音频模型和两个数据集，通过可学习的掩码把嵌入空间切分为属性特定的子空间，用于音乐相似性检索。研究测试了不同掩码与池化方案下新增可训练参数的适配效果。
结果显示，极小的适配容量就足以让事后门控同时提升属性解耦程度和整体检索性能，而池化、掩码方式与适配深度的选择带来的额外收益有限。这表明无需微调整个骨干，仅在冻结嵌入上加少量参数即可实现细粒度属性控制。传统 CSN 需要从头训练骨干模型，成本高；该方法保留了基础模型的通用音乐相似性表征，同时解决了潜在空间纠缠导致的可解释性与属性控制受限问题。
Exploring Post-Hoc Disentanglement of Audio Foundation Models for Music Similarity
Audio foundation models like CLAP, MuQ, and MERT excel at capturing general musical similarity. On the other hand, their entangled latent spaces limit interpretability and fine-grained attribute control. Conditional Similarity Networks (CSN) address this by pairing a backbone representation model with learned masks that gate the embedding space into attribute-specific subspaces. However, existing methods train the backbone from scratch, rather than leveraging pretrained foundation models. We instead present a systematic study of post-hoc disentanglement, adapting CSN to gate frozen pretrained embeddings. We evaluate adding trainable parameters to adapt the embedding under different masking and pooling regimes across two datasets. Our findings show that minimal adaptation capacity is sufficient for post-hoc gating to improve both attribute-specific disentanglement and overall retrieval performance, while pooling, masking, and depth choices yield limited additional gains.
This report is categorized under AITOP's "Papers" desk. Key entities and technical topics under continuous tracking include CLAP, MuQ, MERT, 音乐检索.
Source attribution: curated from arXiv cs.LG (@Lucas de Souza Brandão, Gabriel da Costa Merlin, Diego Furtado Silva).