End-to-End Multi-Task Learning for Adjustable Joint Noise Reduction and Hearing Loss Compensation
A single neural network trained end-to-end outperforms separate models and matches traditional prescriptions.
A research team from the Technical University of Denmark (DTU) has published a breakthrough paper on arXiv detailing a new AI framework for advanced hearing aid processing. Their system uses a single deep neural network (DNN) trained with a multi-task learning approach to handle both noise reduction (NR) and hearing loss compensation (HLC) simultaneously. A key innovation is the use of an inherently differentiable auditory model, allowing for true end-to-end optimization of the entire system, unlike previous methods that required separate emulator models.
During inference, the DNN predicts two separate time-frequency masks—one for NR and one for HLC. Users can independently adjust the intensity of each effect by simply exponentiating these masks before they are combined. This provides unprecedented user control. Furthermore, the system is personalized by feeding the listener's specific audiogram directly into the network as an input, eliminating the need for costly and time-consuming retraining for each individual user.
The results demonstrate significant advantages. The single jointly-trained network not only allows for independent adjustment of noise reduction and amplification but also achieves superior objective performance metrics compared to optimizing for a single objective. Crucially, it outperforms a cascade of two separate DNNs (one trained solely for NR, another for HLC) and delivers HLC performance competitive with traditional hearing-aid prescription methods. This marks the first study to successfully train one DNN for both core hearing aid functions across a wide spectrum of listener profiles using an auditory model.
- Uses a single DNN with multi-task learning for joint noise reduction and hearing loss compensation, outperforming separate cascaded models.
- Features adjustable user controls via exponentiation of predicted time-frequency masks, allowing independent tuning of NR and HLC levels.
- Accepts an audiogram as input for instant listener personalization without requiring model retraining, enabling scalable customization.
Why It Matters
This AI architecture paves the way for smarter, more adaptable, and highly personalized hearing aids that users can fine-tune in real-time.