Semantic-Aware Bias Estimation Framework for Fairness in Large Audio Language Models
A new research paper proposes a semantic-aware mixed-effects regression framework to evaluate fairness in Large Audio Language Models (LALMs), addressing confounding factors such as semantic variation and speaker-specific characteristics. The framework incorporates sentence-level semantic embeddings from the LALM itself as covariates and models speaker identity as a random effect, enabling control over variation as perceived by the model. Experiments on simulated and real-world data demonstrate the approach's effectiveness in mitigating misleading conclusions about model bias. The paper is available on arXiv (2608.13624) and was announced as a cross-type submission.
Key facts
- The paper proposes a semantic-aware mixed-effects regression framework for fairness evaluation in LALMs.
- The framework accounts for confounding factors including semantic variation and speaker-specific characteristics.
- Semantic representations are extracted from the same LALM under evaluation.
- Speaker identity is modeled as a random effect.
- Experiments were conducted on simulated and real-world data.
- The paper is available on arXiv with ID 2608.13624.
- The announcement type is 'cross'.
- The paper addresses fairness concerns in audio understanding tasks such as speech recognition and audio question answering.
Entities
Institutions
- arXiv