주성분 분석법을 이용한 머리전달함수 모형화 기법의 성능 비교

We proposed a new head-related transfer function (HRTF) interpolation method based on splicing. The sound spreads from sound source to listener’s ears, the wave of head-related impulse response (HRIR) clearly indicates time delay which can be symbolized by the sample number of the first obvious peak and the signal filtered by the diffraction and reflection properties of the torso, head, and pinna. Based on this feature and the importance of the time delay in interpolation, we split the measured HRIR into three parts according to the first obvious peak of the wave and the length of the useful signal. First, we used radial basis function neural network to construct regression forecast models using spatial parameters (i.e., azimuth, elevation, and distance) and the HRIRs sample number of the first obvious peak. Then, we used the tetrahedral interpolation with barycentric weights to calculate the second part of the HRIR. Finally, we spliced the forecast information and the useful signal in terms of the sample number of the first obvious peak. For an unknown position in three-dimensional space of near-field, we interpolated its HRIR by the proposed method. The Signal-to-Deviation ratio (SDR) increased 5 dB compared with the tetrahedron interpolation in the time domain.

A complex-valued head-related transfer function (HRTF) can be represented as a real-valued head-related impulse response (HRIR). The interaural time and level cues of HRIRs are extracted to derive the binaural model and also to normalize each measured HRIR. Using the Karhunen-Loeve expansion, normalized HRIRs are modeled as a weighted combination of a set of basis functions in a low-dimensional subspace. The basis functions and the space samples of the weights are obtained from the measured HRIR. A simple linear interpolation algorithm is employed to obtain the modeled binaural HRIRs. The modeled HRIRs are nearly identical to the measured HRIRs from an anesthetized live cat. Typical mean-square errors and cross-correlation coefficients between the 1816 measured and modeled HRIRs are 1% and 0.99, respectively. The real-valued operations and linear interpolating in the model are very effective for speeding up the model computation in real-time implementation. This approach has made it possible to simulate real free-field signals at the two eardrums of a cat via earphones and to study the neuronal responses to such a virtual acoustic space (VAR).

3D audio real-time rendering is of great importance for virtual reality (VR) application, especially on mobile devices. However, the limited computational power makes it hard to implement fast generation of spatial sound using head-related transfer function (HRTF). To accelerate the procedure of head-related impulse response (HRIR) convolution, we propose an efficient rendering method using the parameterized HRTF. The HRIRs are preprocessed into the parameterized HRTFs with a delayed parameter, so as to realize the fast convolutions in 3D audio synthesis. We found our method could reduce the calculation time of 3D audio rendering significantly. The results were 0.9 times and 0.68 times on average the play duration of the MPEG standard sequences, with Qualcomm Snapdragon 821 and Samsung Exynos 7420. And the subjective test showed that the real-time rendered sounds had the similar audio quality and location perception to the 3D audio rendered with the raw HRIRs.

Purpose – The purpose of this paper is to present a system for customized binaural audio delivery based on the extraction of relevant features from a 2-D representation of the listener’s pinna. Design/methodology/approach – The most significant pinna contours are extracted by means of multi-flash imaging, and they provide values for the parameters of a structural head-related transfer function (HRTF) model. The HRTF model spatializes a given sound file according to the listener’s head orientation, tracked by sensor-equipped headphones, with respect to the virtual sound source. Findings – A preliminary localization test shows that the model is able to statically render the elevation of a virtual sound source better than non-individual HRTFs. Research limitations/implications – Results encourage a deeper analysis of the psychoacoustic impact that the individualized HRTF model has on perceived elevation of virtual sound sources. Practical implications – The model has low complexity and is suitable for implementation on mobile devices. The resulting hardware/software package will hopefully allow an easy and low-tech fruition of custom spatial audio to any user. Originality/value – The authors show that custom binaural audio can be successfully deployed without the need of cumbersome subjective measurements.

Head-related transfer function (HRTF) plays an important role in the construction of 3D auditory display. This paper presents an individual HRTF modeling method using deep neural networks based on spatial principal component analysis. The HRTFs are represented by a small set of spatial principal components combined with frequency and individual-dependent weights. By estimating the spatial principal components using deep neural networks and mapping the corresponding weights to a quantity of anthropometric parameters, we predict individual HRTFs in arbitrary spatial directions. The objective and subjective experiments evaluate the HRTFs generated by the proposed method, the principal component analysis (PCA) method, and the generic method. The results show that the HRTFs generated by the proposed method and PCA method perform better than the generic method. For most frequencies the spectral distortion of the proposed method is significantly smaller than the PCA method in the high frequencies but significantly larger in the low frequencies. The evaluation of the localization model shows the PCA method is better than the proposed method. The subjective localization experiments show that the PCA and the proposed methods have similar performances in most conditions. Both the objective and subjective experiments show that the proposed method can predict HRTFs in arbitrary spatial directions.

Head-Related Transfer Function (HRTF) plays a critical role in how the auditory system perceives spatial information. The spectral cues embedded in HRTF are vital for accurately determining the elevation of sound sources. In existing approaches, deep neural networks (DNNs) have been utilized to predict the magnitude spectra of HRTF from images of the pinna, typically employing the HRTF log-magnitude as the output during training. However, HRTF encompasses the acoustic characteristics of both the head and torso, exhibiting direction-dependent patterns that pose challenges in reconstructing its spectral cues. To address this complexity, we propose an innovative method for HRTF individualization. Our model uses Pinna-Related Transfer Function (PRTF) as the output during training, which helps alleviate the impact of sound reflections from the head and torso in the head-related impulse response (HRIR). Our experimental findings, based on an HRTF dataset, illustrate that our proposed model excels in reconstructing the first and second spectral cues. Furthermore, it outperforms previous deep learning models in terms of log spectral distortion (LSD).

Head-related transfer functions (HRTFs) vary with both frequency and source position. The near-field HRTFs with source distance less than 1.0 m are particularly complicated due to their distance dependence. Principal component analysis (PCA), which is conventionally carried out in the frequency or time domain, has been widely applied to reduce the dimensionality of far-field HRTF data with source distance greater than 1.0 m. The present work firstly extends the conventional PCA, and then proposes a spatial PCA method in the spatial domain rather than in the frequency or time domain to reduce the dimensionality of near-field HRTF data. An illustrative case indicates that near-field HRTF magnitudes at 9 distances with 493 directions for each distance can be approximately represented by the weighted sum of 15 spectral shape basis functions using the conventional PCA or the weighted sum of 15 spatial basis functions using the spatial PCA. Both representations account for more than 98 % energy variation of the original data, and reduce the dimensionality of the original data to about a quarter. The proposed method is also applicable to the head-related impulse responses in the time domain. Furthermore, the spatial PCA scheme is potentially applicable to simplify near-field HRTF measurement.

Human beings have a remarkable ability to determine the direction of arrival of a sound and to separate sounds of interest. Replicating this ability is a challenging problem in audio signal processing. In this paper we present a model for the head related transfer function (HRTF) developed with the localisation objective in mind. This is achieved by splitting the 3D localisation cues in terms of two functions which can be independently evaluated. We illustrate the theory for calculating these functions and validate the results against actual HRTF data. We find the model to be a close match for a significant number of potential source locations.

Known errors exist in loudspeaker array processing techniques, often degrading source localization and timbre. The goal of the present study was to use virtual loudspeaker arrays to investigate how treatment of the interaural time delay (ITD) cue from each loudspeaker impacts these errors. Virtual loudspeaker arrays rendered over headphones using head-related impulse responses (HRIRs) allow flexible control of array size. Here, three HRIR delay treatment strategies were evaluated using minimum-phase loudspeaker HRIRs: reapplying the original HRIR delays, applying the relative ITD to the contralateral ear, or separately applying the HRIR delays prior to virtual array processing. Seven array sizes were simulated, and panning techniques were used to estimate HRIRs from 3000 directions using higher-order Ambisonics, vector-base amplitude panning, and the closest loudspeaker technique. Compared to a traditional, physical array, the prior HRIR delay treatment strategy produced similar errors with a 95% reduction in the required array size. When compared to direct spherical harmonic (SH) fitting of head-related transfer functions (HRTFs), the prior delays strategy reduced errors in reconstruction accuracy of timbral and directional psychoacoustic cues. This result suggests that delay optimization can greatly reduce the number of virtual loudspeakers required for accurate rendering of acoustic scenes without SH-based HRTF representation.

Currently, to obtain maximum fidelity 3D audio, an intended listener is required to undergo time consuming measurements using highly specialized and expensive equipment. Customizable Head-Related Impulse Responses (HRIRs) would remove this limitation. This paper reports our progress in the first stage of the development of customizable HRIRs. Our approach is to develop compact functional models that could be equivalent to empirically measured HRIRs but require a much smaller number of parameters, which could eventually be derived from the anatomical characteristics of a prospective listener. For this first step, HRIRs must be decomposed into multiple delayed and scaled damped sinusoids which, in turn, reveal the parameters (delay and magnitude) necessary to create an instance of the structural model equivalent to the HRIR under analysis. Previously this type of HRIR decomposition has been accomplished through an exhaustive search of the model parameters. A new method that approaches the decomposition simultaneously in the frequency (Z) and time domains is reported here.

This paper introduces a continuous functional model for head-related transfer functions (HRTFs) in the horizontal auditory scene. The approach uses a separable representation consisting of a Fourier-Bessel series expansion for the spectral components and a conventional Fourier series expansion for the spatial components. Being independent of the data, these two sets of basis functions remain unchanged for all subjects and measurement setups. Hence, the model can transform an individualized HRTF to a subject specific set of coefficients. A continuous functional model is also developed in the time domain. We show the efficient model performance in approximating experimental measurements by using the HRTF measurements from a KEMAR manikin and the synthetic data from the spherical head model. The statistical results are determined from a 50-subject HRTF data set. We also corroborate the predictive capability of the proposed model. The model has near optimal performance, which can be ascertained by comparison with the standard principle component analysis (PCA) and discrete Karhunen-Loeve expansion (KLE) methods at the measurement points and for a given number of parameters.

This paper provides a small set of essential parameters for a personalized and effective real-time auralization with headphones. An image-guided procedure with two 2D images of the user’s head guides the mixed structural modeling of head-related transfer function (HRTF), combining a spherical head model with ear displacement with the HRTF high-frequency magnitude selected from a database according to ear anthropometry. Room acoustics phenomena are simplified following the scattering delay network (SDN) approach which allows an accurate spatialization of first order reflections. Finally, statically significant improvements in localization performances within a virtual reality (VR) test allow to identify some benefits of the proposed customized auralization model compared to the widely used higher-order ambisonics (HOA) rendering with generic HRTFs.

In this paper an approach to design controllers based on principal components analysis (PCA) models is presented. Closed-loop control can be formulated and implemented within the reduced space defined by a PCA model. This PCA controller, results in an integral controller, which can be used as an inferential controller when a measurement of a primary variable is not available. The main contributions of the paper are: a) the incorporation of an adjustable gain on the classical PCA controller; b) the proposal of a set of tuning rules; c) the performance evaluation of this new controller, in nominal operation conditions and in faulty situations. Some experimental results, obtained with the three tank benchmark (European COSY project), are presented.

Head-related transfer function (HRTF) interpolation plays an important role for implementation of 3D sound system because it can not only reduce the number of measurements for HRTFs, but also reduce the data of HRTFs for seamless binaural synthesis. This paper addresses the problem of accurately realizing the interpolation of HRTF for synthesis of virtual auditory space, and proposes a HRTF interpolation method based on principal component analysis. Firstly, the HRTF is decomposed into principal components and corresponding principal component weights, where principal components are direction-independent and principal component weights are direction- dependent; then the directional variation of the principal component weight is multivariate polynomial fitted with a bivariate function of two spatial angulus (azimuth and elevation). Moreover, a sphere- partitioning optimization scheme is employed to improve the approximation precision. Experiment results demonstrate that HRTFs in the entire sphere surface can be interpolated by the proposed method with small distortion, and the proposed method performs better than conventional methods. Therefore the proposed method gives a promising way for HRTF interpolation.

This paper presents a deep-learning technique for modeling Head-related Transfer Function (HRTF) using a large data-set from IRCAM [7] of 49 subjects with 115 directions/subject. We use a log-spectral distortion metric that we have used to optimize the performance of the deep-learning model. The presented model first uses a sparse autoencoder (AE) for capturing the salient representation (peaks, and notches for elevation) from the user-captured HRTFs, in its latent representation, and then uses a generalized regression neural network (GRNN) to approximate the latent representation of the corresponding HRTFs at arbitrary angles. The sparse AE then decodes the GRNN output to reconstruct the HRTF. The proposed approach is shown to outperform, objectively using the log-spectral distortion, the state-of-the-art linear-least-squares optimum solution (viz., derived from Principal Component Analysis) to perform better generalization over the HRTFs in the IRCAM dataset. Towards this, Bayesian optimization has been performed to tune the hyper-parameters of the sparse AE and the GRNN.