Subspace-based HRTF Synthesis from Sparse Data: A joint PCA and ML-based Approach

摘要

Head-related transfer functions (HRTF) are used for creating the perception of a virtual sound source at an arbitrary azimuth-elevation. Publicly available databases use a subset of these directions due to physical constraints (viz., loudspeakers for generating the stimuli not being point-sources) and the time required to acquire and deconvolve responses for a large number of spatial directions. In this paper, we present a subspace-based technique for reconstructing HRTFs at arbitrary directions for the IRCAM-Listen HRTF database, which comprises a set of HRTFs sampled every 15 deg along the azimuth direction. The presented technique includes first augmenting the sparse IRCAM dataset using the concept of auditory localization blur, then deriving a set of P=6 principal components, using PCA for the original and augmented HRTFs, and then training a neural network (ANN) with these directional principal components. The reconstruction of HRTF corresponding to an arbitrary direction is achieved by post-multiplying the ANN output, comprising the estimated six principal components, with a frequency weighting matrix. The advantage of using a subspace approach, involving only 6 principal components, is to obtain a low complexity HRTF synthesis ANN-based model as compared to training an ANN model to output an HRTF over all frequencies. Objective results demonstrate a reasonable interpolation with the presented approach.