Institute of Information Science, Academia Sinica

This talk presents HAAQI-Net, a novel non-intrusive deep learning-based model for music audio quality assessment tailored for hearing aid users. Unlike traditional intrusive methods that require reference signals, HAAQI-Net leverages a Bidirectional Long Short-Term Memory (BLSTM) network with attention mechanisms and features extracted from the pre-trained BEATs model to predict HAAQI scores directly from music audio clips and hearing loss patterns. Experimental results demonstrate its effectiveness, achieving a high correlation with traditional metrics while significantly reducing inference time. Additionally, a knowledge distillation strategy reduces computational complexity by 75.85%, maintaining strong performance while accelerating inference. Beyond objective assessment, HAAQI-Net is fine-tuned to predict subjective Mean Opinion Scores (MOS), improving alignment with human perception. The model’s robustness under varying Sound Pressure Levels (SPL) is also explored, revealing optimal performance at a reference SPL of 65 dB, with accuracy gradually decreasing as SPL deviated from this point. These advancements position HAAQI-Net as a scalable and efficient solution for music quality assessment in hearing aid applications.