Abstract:

In spatial hearing, mass data of head-related transfer function (HRTF) is a factor that greatly influences the synthesis of virtual sounds. To reduce the data used, we propose a spatial hearing reconstruction method based on locally linear embedding (LLE). Using LLE, high dimensionality is mapped to a lower dimensional dataset suitable for regressive analysis and classification. To classify by an unsupervised cluster method, a representative HRTF is extracted from all HRTFs. Other HRTFs can be reconstructed in spatial hearing from the representative HRTF with modified interpolation. Compared to the principal component analysis (PCA), the data with reduced dimension obtained by using LLE preserve more perceptive information. Relations among the HRTF data can be found, with which modified interpolation can be obtained and HRTF reconstruction error can be reduced. Simulation results show that the proposed method effectively reduces HRTF data. It is useful to improve efficiency of synthesis for virtual sound source in practical applications.

Key words: head-related transfer function (HRTF); locally linear embedding (LLE); manifold; spatial hearing reconstruction