A localization algorithm based on head-related transfer functions

摘要

A localization algorithm based on head-related transfer functions (HRFTs) is introduced. The algorithm minimizes the Euclidean distance to direction-dependent binaural signal spaces, in order to find an estimator of the actual sound source position. Numerical experiments with different target sounds (clicks, white noise, speech, rustling of leaves) and two signal-to-noise ratios (10 dB and 0 dB), where the masker is non-spatialized white noise, show that this projection algorithm outperforms template matching and the cross-channel algorithm in nearly all experimental conditions. In particular, the algorithm is robust to changes of the emitted signal's phase spectrum, unlike template matching. For white noise maskers, it is possible to compute SNR dependent estimates for the error probability in the task of discriminating two directions, based on the associated HRTFs. We present simulations that demonstrate the precision of the estimates. We show how these probabilities can be employed as a means to mathematically analyze HRTFs, in particular with the aim of predicting localization performance from the HRTF data set.