HRTF Database Research Articles

Employing deep learning method to predict global head-related transfer functions from scanned head geometry

We propose an HRTF personalization method in which a Convolutional Neural Network (CNN) is employed to learn subjects' HRTFs from the scanned geometry of their heads. The trained model can then be employed to predict the global HRTF set (for all directions) from the subject's head scan data alone. In our trial, the HUTUBS HRTF database was used as the training set. A truncated spherical harmonic expansion of head scan data was used to preserve the boundary shape features that are important in the acoustic scattering process. Each HRTF for a given subject also was represented by a truncated spherical harmonic expansion. The SHT coefficients of the scanned head geometry and the HRTFs serve as training data for a CNN that subsequently can be used to predict HRTFs from geometric scan data. A leave-one-out validation with log-spectral distortion (LSD) metric was used for evaluation. The results show a decent LSD level at both spatial & temporal dimensions compared to the ground truth, and have lower LSD than the finite element acoustic simulation of HRTFs that the database provides. In continuing work, we are validating the prediction results in listener tests.

Deployment of models of HRIRs using PCA in digital signal processing board

Three-dimensional sound effect has been around us when we are playing video games and watching movies in theatres. This happens because many researches have been developed to deploy head-related impulse responses (HRIRs) for 3-D sound reproduction. In reproducing 3-D sound, monaural sound is convolved with a pair of HRIRs. In this research, PKU-IOA HRTF Database has been used. An original HRIR consists of 1024 samples with 65,536 Hz sampling rates. Pre-processing is performed on the HRIRs, so that we obtain five data types including the original ones, i.e. original HRIRs (type 1); original HRIRs chopped to first 512 samples (type 2); and 349 samples (type 3); HRIRs with 256 samples and 48 kHz sampling rate (type 4) that originating from original HRIRs with 349 samples that are down-sampled to 48 kHz; and the last data type is minimum phase HRIRs with first 200 samples that are reconstructed from original HRIRs (type 5). Each data type is then modelled using principal components analysis (PCA). The averages of percentage errors are respectively 24.259%, 24.244%, 24.242%, 21.080%, and 5.422% for type 1 to type 5. The best-resulted model is minimum phase HRIRs (type 5); these HRIRs are then reconstructed and deployed in Digital Signal Processor (DSP) Board TMS320C5535 to reproduce 3-D sound from monaural one.

A New Selection Method of Anthropometric Parameters in Individualizing HRIR

A trend issue in modeling head-related impulse responses (HRIRs) is how to individualize HRIRs models that are convenient for a particular listener. The objective of this research is to show a robust selection method of eight anthropometric parameters out of all 27 parameters defined in CIPIC HRTF Database. The proposed selection method is systematically and scientifically acceptable, compared to ‘trial and error’ method in selecting the parameters. The selected anthropometric parameters of a given listener were applied in establishing multiple linear regression models in order to individualize his / her HRIRs. We modeled the entire minimum phase HRIRs in horizontal plane of 35 subjects using principal components analysis (PCA). The individual minimum phase HRIRs can be estimated adequately by a linear combination of ten orthonormal basis functions.

Analysis of Korean head-related transfer function

It is necessary to construct head-related transfer function database for rendering and studying three dimensional audio. For this reason, many research groups have tried to develop a HRTF measurement system and to construct a HRTF database for their research. Even though there are various HRTF databases, there is no database with anthropometry in public domain aimed at Koreans even if the HRTFs vary based on physical shapes of subjects. Because Koreans hear three dimensional sound rendered by HRTF database based on Caucasians, performance of three dimensional sound might be hindered. To verify this possibility and remedy the drawbacks of established HRTF database, construction of new HRTF database aimed at Korean is needed. For constructing HRTF database, new HRTF measuring system using sine sweep signal was developed and the HRTFs for 10 subjects at 49 different elevation and 36 different azimuths at 5angular increments were measured. By using measured HRTFs, the HRTFs aimed at Koreans were compared with CIPIC HRTF database and analyzed.

Analysis on perceptual sensitivity to head-related impulse responses in the median plane

This study deals with the perceptual sensitivity to Head-Related Impulse Responses (HRIRs) in the median plane based on a series of subjective listening tests using a pair of headphones. First, the non-individualized HRIRs were modeled from 12 principal components (PCs) extracted from Principal Components Analysis (PCA) of the CIPIC HRTF database. The Just Noticeable Difference (JND) in weight of PCs (PCWs) at each elevation was estimated. It was not observed the common elevation-dependent tendency or PCW-dependent tendency of JND in PCWs across the five subjects who participated in the tests, and the inter-subject variation of JND in PCWs was large. The JND in HRIRs can be estimated indirectly from the JND in PCWs because the HRIRs can be represented by a linear summation of the PCs weighted by PCWs. The common elevation-dependent tendency of JND in Directional Impulse Responses (DIRs), which are the mean-subtracted HRIRs, across the five subjects can be found. The change in PCWs does not seem to contribute to our perception of sound source characteristics; however, the resulting change in HRIRs due to the change in PCWs seems to contribute. The subjects showed larger JND in DIRs in the frontal region than in the rear region. This means that our perception of sound source characteristics is more sensitive for frontal sources than rear sources.

보편적인 기저함수를 이용한 중앙면상의 머리전달함수 모델링

A principal components analysis (PCA) of the median-plane head-related impulse responses (HRIRs) in the CIPIC HRTF database reveals that the individual HRIRs in the median plane can be adequately reconstructed by a linear combination of 12 orthonormal basis functions. These basis functions can be used to model arbitrary median-plane HRIRs, which are not included in the process to obtain the basis functions. Memory size can be reduced up to 5-fold depending on the number of HRIRs to be modeled. To clarify whether these basis functions can be used to model other set of arbitrary median plane HRIRs, a numerical error analysis for modeling and a series of subjective listening tests were carried out using the measured and modeled HRIRs. The results showed that the set of individual HRIRs in the median plane, which were measured in our lab using different measurement conditions, techniques, and source positions, can be modeled with reasonable accuracy. All subjects, involved in the subjective listening test, reported not only the accurate vertical perception but also the front-back discrimination with the modeled HRIRs based on 12 basis functions.

Interpretations on principal components analysis of head-related impulse responses in the median plane

A principal components analysis of the median-plane head-related impulse responses (HRIRs) in the CIPIC HRTF database reveals that the individual HRIRs can be reconstructed by a linear combination of 12 principal components (PCs) within 5% of error in the least-squares sense. The PCs include the intersubject and interelevation variations in the median-plane HRIRs. Each PC provides sound cues for the front-back discrimination and/or the vertical perception. There exist common systematic elevation dependencies in the weights of lower-numbered PCs which contribute to the pinna/head diffractions, whereas the elevation dependencies in the weights of higher-numbered PCs are different from subject to subject.

주성분 분석을 이용한 HRIR 맞춤 기법

A principal components analysis of the entire median HRIRs in the CIPIC HRTF database reveals that the individual HRIRs can be adequately reconstructed by a linear combination of several orthonormal basis functions. The basis functions represent the inter-individual and inter-elevation variations in median HRIRs. There exist elevation-dependent tendencies in the weights of basis functions, and the basis functions can be ordered according to the magnitude of standard deviation of the weights at each elevation. We propose a HRIR customization method via tuning of the weights of 3 dominant basis functions corresponding to the 3 largest standard deviations at each elevation. Subjective listening test results show that both front-back reversal and vertical perception can be improved with the customized HRIRs.