A Sound Image Reproduction Model Based on Personalized Weight Vectors

摘要

Many perceptual models for audio reconstruction have been proposed to create the virtual sound, but the direction of the virtual sound maybe deviate from the desired direction due to the distortion of binaural cues. In this paper, a binaural cues' equation for real sound and virtual one reproduced by dual loudspeakers is established to derive weight vectors based on the head-related transfer function (HRTF). After being filtered by the weight vectors, sound signals emitted from the loudspeakers can deliver an accurate spatial impression to the listener. However, the HRTFs change with listeners, by which the weight vectors calculated also vary from person to person. Therefore, a radial basis function neural network (RBFNN) is designed to personalize weight vectors for each specific listener. Compared with the three methods including vector base amplitude panning (VBAP), the HRTF-based panning (HP) and the band-based panning (BP), the method in this paper can reproduce binaural cues more accurately, and subjective test also indicates that there is no significant difference in perception between real sound and virtual sound based on the proposed methods.