aiqbench analysis
This paper identifies a specific failure mode in speech emotion recognition and audio LLMs: when vocal tone contradicts the words spoken, models lean heavily on the text. It contributes a new adversarial benchmark (ASPIRE), two evaluation metrics, and an architecture (ACR-Net) that uses contrastive decoupling to separate the two signals. The results are reported on the authors' own benchmark and metrics, so readers should check how well the gains transfer to other datasets and real-world conditions before drawing broad conclusions.
ACR-Net: Mitigating Semantic Dominance via Contrastive Acoustic-Semantic Decoupling
arXiv:2610.20114v1 Announce Type: cross Abstract: Speech Emotion Recognition (SER) models and Audio LLMs fail when vocal tone contradicts textual semantics. Extensive training on congruent data engenders a rigid semantic prior, causing "Semantic Dominance", a severe text bias during cross-modal conflicts. We introduce the ASPIRE benchmark of adversarial audio-text pairs across four conflict types, alongside two metrics: Semantic Overconfidence Penalty (SOP) and Latent Decoupling Degree (LDD). Unlike fusion methods that merely re-weight collapsed features, our Acoustic Conflict Resolution Network (ACR-Net) uses Cross-Modal Attention and Contrastive Decoupling Loss to disentangle contradictory representations into orthogonal latent spaces. Experiments show existing models fail under polarity conflicts due to extreme semantic overconfidence. Conversely, ACR-Net preserves acoustic fidelity, maintaining high LDD to drive SOP near zero, achieving superior Acoustic Accuracy (ACC) in both adversarial and standard scenarios.
Abstract:Speech Emotion Recognition (SER) models and Audio LLMs fail when vocal tone contradicts textual semantics. Extensive training on congruent data engenders a rigid semantic prior, causing "Semantic Dominance", a severe text bias during cross-modal conflicts. We introduce the ASPIRE benchmark of adversarial audio-text pairs across four conflict types, alongside two metrics: Semantic Overconfidence Penalty (SOP) and Latent Decoupling Degree (LDD). Unlike fusion methods that merely re-weight collapsed features, our Acoustic Conflict Resolution Network (ACR-Net) uses Cross-Modal Attention and Contrastive Decoupling Loss to disentangle contradictory representations into orthogonal latent spaces. Experiments show existing models fail under polarity conflicts due to extreme semantic overconfidence. Conversely, ACR-Net preserves acoustic fidelity, maintaining high LDD to drive SOP near zero, achieving superior Acoustic Accuracy (ACC) in both adversarial and standard scenarios.
5 pages, 1 figure, 4 tables. Published in Interspeech 2026
Sound (cs.SD); Artificial Intelligence (cs.AI)
arXiv:2610.20114 [cs.SD]
(or arXiv:2610.20114v1 [cs.SD] for this version)
https://doi.org/10.48550/arXiv.2610.20114 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
Proc. Interspeech 2026, pp. 1207-1211 (2026)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 238 words
Related coverage
DeShortcut-Align: Decoupling Spurious Shortcuts for Robust Safety Alignment in Large Reasoning Models
arXiv · Sep 29, 2026
Multimodal Graph Negative Learning
arXiv · Jun 17, 2026
DiTAR+: Dual Optimization for Robust Autoregressive Diffusion Speech Synthesis
arXiv · Sep 16, 2026
SemMSA: Latent Semantic-Aided Robust Multimodal Sentiment Analysis with Incomplete Data
arXiv · Sep 25, 2026