Reduced representations of HRTF datasets: A discriminant analysis approach

This paper discusses reduced representations of HRTF datasets, fully descriptive of one's personalized properties. The data reduction is achieved through elimination of the least discriminative binaural-filter pairs from a set. For this purpose Linear Discriminant Analysis (LDA) was applied on the Music and Audio Research Laboratory (MARL) database of repeated HRTF measurements, which resulted in 67% data reduction. The effectiveness of the sparse HRTF mapping is assessed by way of the performance of a database matching system, followed by a subjective evaluation study. The results indicate that participants have demonstrated strong preference towards the selected HRTF sets, in contrast to the generic KEMAR set and the least similar selections from the repository.