Modeling individual HRTF tensor using high-order partial least squares

Abstract

A tensor is used to describe head-related transfer functions (HRTFs) depending on frequencies, sound directions, and anthropometric parameters. It keeps the multi-dimensional structure of measured HRTFs. To construct a multi-linear HRTF personalization model, an individual core tensor is extracted from the original HRTFs using high-order singular value decomposition (HOSVD). The individual core tensor in lower-dimensional space acts as the output of the multi-linear model. Some key anthropometric parameters as the inputs of the model are selected by Laplacian scores and correlation analyses between all the measured parameters and the individual core tensor. Then, the multi-linear regression model is constructed by high-order partial least squares (HOPLS), aiming to seek a joint subspace approximation for both the selected parameters and the individual core tensor. The numbers of latent variables and loadings are used to control the complexity of the model and prevent overfitting feasibly. Compared with the partial least squares regression (PLSR) method, objective simulations demonstrate the better performance for predicting individual HRTFs especially for the sound directions ipsilateral to the concerned ear. The subjective listening tests show that the predicted individual HRTFs are approximate to the measured HRTFs for the sound localization.

Manifold learning for spatial audio rendering

“…Although our results are slightly better than [174], they are lower than [172], which is -according to our research-the smallest score reported among studies using MSD. Finally, in general, tensor-based approaches [176,175,177,178] perform better than PCA-based methods, reaching their best performance at 3.5 dB in the frequency band 50 Hz -8 kHz [178], which is comparable to our results in the frequency band 0.2 -8 kHz (2.9 dB, σ = 0.0735).…”

“…A single linear regression model is used for mapping anthropometric features to a five-dimensional vector obtained by means of N-mode Singular Value Decomposition (N-mode SVD) and which represents the subject mode in the tensor. A similar tensor-based approach is used in [175] and [177], but to construct the regression model, they employed an ANN and high-order partial least squares, respectively.…”

“…Finally, note that, unlike previous works [167,169,170,171,172,174], we applied dimensionality reduction only once, over the entire dataset, for HRTFs of all subjects, directions and ears. This way, as tensor-based approaches [175,176,177,178] do, we tried to preserve the multi-factor (i.e. frequency, direction and subject) nature of HRTFs.…”

Current Use and Future Perspectives of Spatial Audio Technologies in Electronic Travel Aids

“…Algazi et al [93] suggested using a number of one-dimensional anthropometric measurements for HRTF fitting through regression methods or other machine learning techniques. This approach was recently pursued in a number of studies [126][127][128][129] investigating the correspondence between anthropometric parameters and HRTF shape. When suitable processing is performed on HRTFs, clear relations with anthropometry emerge.…”

Prediction model of sinoatrial node field potential using high order partial least squares

“…In fact, most of the experimental data describe the relations between multi-way input and multi-way output which are deemed as tensors. In order to adjust this kind of data, a new prediction model, higher order partial least squares (HOPLS), was proposed [7,8]. HOPLS is highly suitable for building a prediction model which has tensor-input and tensor-output.…”