Background: Head-related transfer functions (HRTFs) contain time and level cues and may be utilized in automatic algorithms to identify locations of sound, a desirable feature for next-generation hearing devices. Due to substantial variability in individual head sizes and ear acoustics, individualized HRTFs are expected to provide the best localization results. However, acquiring individualized HRTFs for each user is time-consuming. Methods: This study constructed three binaural and/or monaural algorithms suitable for hearing devices. A linear classifier was trained on HRTF databases from a subset of subjects and used to predict sound locations for other individuals to evaluate cross-subject variability. Results: Using the CIPIC Database, a “two-step” method achieved a horizontal localization error of 1.0° and a vertical error of 30.4° sequentially. With the 3D3A Database, the horizontal and vertical errors were 5.6° and 36.5°, respectively. Both datasets yielded improved accuracy when frontal and rear hemifields were simulated separately, with trends remaining consistent across databases. When subjects were grouped by gender, classifiers trained on women’s HRTFs performed well in predicting men’s localization, whereas classifiers trained on men’s HRTFs resulted in significantly larger errors. Conclusions: These findings offer insights into the localization cues embedded in HRTFs and demonstrate the influences of inter-subject variability for spatial hearing devices.
Orr et al. (Fri,) studied this question.