Yang, Yang, Liang, Ruiyu, Ni, Ye, Zou, Cairong, Sun, Song, Yan, Jing, Zhou, Wei, Ding, Weiping and Hao, Xiaoshuai
2027.
BiSQAFusion: Multi-level fusion for personalized binaural speech quality assessment in hearing aids.
Information Fusion: An International Journal on Multi-Sensor, Multi-Source Information Fusion
139
(Part A)
, 104765.
10.1016/j.inffus.2026.104765
|
Preview |
PDF
- Accepted Post-Print Version
Available under License Creative Commons Attribution. Download (12MB) | Preview |
Abstract
Accurate speech quality assessment is essential for optimizing hearing aid (HA) algorithms, enabling personalized fitting, and improving user satisfaction. Existing HA assessment models, however, are largely limited to monaural processing, lack in-depth cross-modal fusion between acoustic features and audiogram priors, and exhibit insufficient robustness across diverse HA configurations. To address these challenges, we propose BiSQAFusion, a multi-level fusion network that jointly integrates multi-layer self-supervised speech representations, individual audiogram priors, interaural interaction cues, and HA system identifiers for personalized non-intrusive binaural HA speech quality assessment. At its core, a hierarchical two-stage acoustic-audiogram fusion module is designed to adaptively inject personalized audiogram priors into dynamic acoustic features by coupling gated feature-wise linear modulation with a deep CNN-BiLSTM backbone. To overcome the inherent limitations of monaural assessment, a lightweight binaural fusion module is proposed to explicitly decouple interaural coherence-, asymmetry-, and correlation-like cues via parameter-free algebraic operations. Furthermore, systematic prediction biases arising from diverse HA processing configurations are effectively mitigated by a configuration-aware System ID embedding. An auxiliary quality-level module is formulated to impose coarse-grained classification constraints on continuous score prediction via multi-task learning. Extensive experiments on the CPC1 and DNS challenge datasets demonstrate that BiSQAFusion significantly outperforms state-of-the-art non-intrusive baselines in both prediction accuracy and cross-scenario generalization.
| Item Type: | Article |
|---|---|
| Date Type: | Publication |
| Status: | Published |
| Schools: | Schools > Computational & Mathematical Sciences Schools > Computer Science & Informatics |
| Additional Information: | RRS policy applied |
| Publisher: | Elsevier |
| ISSN: | 1566-2535 |
| Date of First Compliant Deposit: | 14 October 2026 |
| Last Modified: | 14 Oct 2026 08:45 |
| URI: | https://orca.cardiff.ac.uk/id/eprint/189605 |
Actions (repository staff only)
 |
Edit Item |
Dimensions
Dimensions