Beyond Genre: Diagnosing Bias in Music Embeddings Using Concept Activation Vectors
2025
Read Beyond Genre: Diagnosing Bias in Music Embeddings Using Concept Activation Vectors (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-24482.txt.
Description
Music models learn genre, but they may quietly learn who is singing as well. Roman B Gebhardt and 2 co-authors use Concept Activation Vectors to probe whether singer traits such as gender and language leak into genre representations. They test four leading music models on a data set balanced to control for genre confounds.
Each model showed its own significant biases, in line with disparities reported in music research and sociology. The team then proposes a fix that edits the concept vectors after training. That manipulation cut the measured biases in their tests. Concept-based tools, they argue, can both diagnose and treat representational bias in music models.
Catalog record
| YEAR | 2025 |
|---|---|
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 33 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2509-24482