An individualization approach for head-related transfer function in arbitrary directions based on deep learning.

Area of Science:

• Acoustics and Signal Processing
• Artificial Intelligence and Machine Learning
• Biomedical Engineering

Background:

• Head-related transfer functions (HRTFs) are crucial for realistic 3D audio rendering.
• Individualizing HRTFs is essential for accurate spatial audio perception.
• Current HRTF individualization methods face challenges in arbitrary direction modeling and spectral accuracy.

Purpose of the Study:

• To develop a deep learning-based approach for HRTF individualization in arbitrary directions.
• To establish a robust relationship between HRTF magnitude spectrum, direction, and anthropometric parameters.
• To improve the spectral accuracy of individualized HRTFs.

Main Methods:

• Utilized a dual-autoencoder architecture combining a variational autoencoder (VAE) and an autoencoder (AE).
• The VAE extracted features from full-space HRTF spectra.
• The AE embedded features from directional and anthropometric parameters.
• A deep neural network (DNN) model was trained to map these features.

Main Results:

• The proposed deep learning model effectively established relationships between HRTF spectra, direction, and anthropometric data.
• Experimental results demonstrated superior performance compared to state-of-the-art methods.
• The method achieved lower spectral distortion, indicating improved HRTF accuracy.

Conclusions:

• The dual-autoencoder deep learning approach offers an effective method for HRTF individualization in arbitrary directions.
• This technique enhances the accuracy of personalized 3D audio rendering.
• The findings suggest a promising direction for future research in spatial audio and virtual acoustics.