Local spectral feature extraction and compaction for HRTFs by nonnegative tensor factorization

Abstract

Head-related transfer functions (HRTFs) depend on frequencies, sound directions, and individuals with high dimensionality and complicated structure. In practical applications, it is difficult to utilize the original HRTFs. Nonnegative matrix factorization (NMF) has been used to reduce the dimension of original HRTFs by vectorization and might lose the information of their natural structure. In this paper, to keep the inherent multi-dimensional structure, a tensor is firstly used to describe HRTFs. Then local spectral features in a low-dimensional space are extracted from HRTFs tensor by nonnegative tensor factorization (NTF). Due to their nonnegativity, the high-dimensional HRTFs tensor can be explained by an additive linear combination of these local spectral features. The simulations demonstrate that NTF achieves higher compression ratio with lower reconstruction error than NMF for HRTFs.