Abstract
Individualized head-related transfer functions (HRTFs) play an important role in accurate localization perception. However, it is a great challenge to efficiently measure continuous HRTFs for each person in full space. In this paper, we propose a parameter-transfer learning method termed PTL to obtain individualized HRTFs based on a small set of HRTF measurements. The key idea behind PTL is to transfer a HRTF generation model from other database to a target individual. To this end, PTL first pretrains a deep neural network (DNN)-based universal model on a large database of HRTFs with the assist of domain knowledge. Domain knowledge is used to generate the input features derived from the solution to sound wave propagation equation at the physical level, and to design the loss function based on the knowledge of objective evaluation criterion. Then, the universal model is transferred to a target individual by adapting the parameters of a hidden layer of DNN with a small set of HRTF measurements. The adaptation layer is determined by experimental verification. We also conduct the objective and subjective experiments, and the results show that the proposed method outperforms the state-of-the-arts methods in terms of LSD and localization accuracy.
Introduction
Spatial audio is a cornerstone of immersive experiences in virtual and augmented reality (VR/AR). To achieve realistic 3D sound, audio must be filtered through Head-Related Transfer Functions (HRTFs), which describe how the anatomy of a listener's head, ears, and torso modifies sound waves. Because these anatomical features are unique to each individual, generic HRTFs often lead to perceptual errors such as front-back confusion or poor elevation localization.
The "gold standard" for individualization is acoustic measurement in an anechoic chamber, but this is a high-resource process involving hundreds of measurements, making it impractical for consumer applications. This paper addresses the "low-resource" challenge by proposing a Parameter-Transfer Learning (PTL) method to generate high-quality individualized HRTFs from a sparse set of measurements.
The PTL Framework
The proposed PTL method utilizes a two-stage deep learning approach to bridge the gap between general acoustic models and individual-specific signatures.
- Universal Model Pre-training: A deep neural network (DNN) is first trained on a large-scale HRTF database (PKU&IOA). To enhance the model's physical relevance, the authors incorporate domain knowledge. Input features are derived from the Spherical Fourier-Bessel (SFB) basis, which provides a mathematical foundation for sound wave propagation. Furthermore, the loss function is based on Log-Spectral Distortion (LSD), a metric that correlates more closely with human auditory perception than standard mean-squared error.
- Model Adaptation (Transfer Learning): The pre-trained universal model is then adapted to a specific individual using a small set of sparse measurements (e.g., 60 samples). The researchers investigated adapting different layers of the network and found that fine-tuning the last hidden layer (PTL-L) yielded the most significant performance gains. This suggests that while early layers learn general spatial features, the final layers are responsible for mapping those features to the specific acoustic "fingerprint" of the user.
Experimental Results
The performance of the PTL method was evaluated using both objective acoustic metrics and subjective listening tests. The objective evaluation utilized the CIPIC database, comparing PTL against existing state-of-the-art linear methods and a non-adapted DNN baseline.
As shown in the table below, the PTL-L variant achieved the lowest Log-Spectral Distortion (LSD) of 4.95 dB and the lowest Root Mean Square Error (RMSE) of 0.072. This represents a significant improvement over previous methods and demonstrates the superiority of the non-linear DNN approach in capturing complex anatomical influences.
Subjective localization tests further validated these findings. Using PTL-L individualized HRTFs, the correct localization rate (CR) improved by approximately 6.24% compared to generic HRTFs, while the front-back confusion rate (FBR) was reduced by 3.98%. These results indicate that the PTL method effectively reduces the "acoustic illusions" typically associated with non-individualized spatial audio.
Conclusion
The Parameter-Transfer Learning (PTL) method offers a robust solution for the low-resource individualization of HRTFs. By combining physics-informed pre-training with targeted layer adaptation, the model achieves high localization accuracy with only a fraction of the measurements required by traditional methods. Future work will focus on optimizing the selection of measurement directions to further reduce the data requirements for personalized 3D audio.
Summary updated