A Survey on Machine Learning Techniques for Head-Related Transfer Function Individualization - Machine Learning Survey

Aug 3, 2026

Machine learning has become essential to binaural synthesis personalization for realistic audio in immersive virtual environments. This survey systematically reviews machine-learning approaches to head-related transfer function (HRTF) individualization. It organizes prior work by the processing steps of the ML workflow, covering datasets, input/output representations, preprocessing, model choices, and evaluation strategies. Beyond classification, it summarizes reported achievements, highlights limitations, and identifies open directions requiring further study across acoustics, audio signal processing, and machine learning research communities.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages27

Reading Time1 h 8 min

AccessFree to download

Introduction

Spatial audio and HRTFs

Motivation for HRTF individualization

Survey organization of ML-based methods

Received 5 August 2024; revised 11 November 2024; accepted 4 December 2024. Date of publication 10 January 2025; date of current version 4 February 2025. The review of this article was arranged by Associate Editor Romain Serizel. Digital Object Identiﬁer 10.1109/OJSP.2025.3528330 A Survey on Machine Learning Techniques for Head-Related Transfer Function Individualization DAVIDE FANTINI , MICHELE GERONAZZO (Senior Member, IEEE), FEDERICO AVANZINI , AND STAVROS NTALAMPIRAS 1Laboratory of Music Informatics (LIM), Department of Computer Science, University of Milan, 20133 Milan, Italy 2Department of Engineering and Management, University of Padua, 35122 Padova, Italy 3Dyson School of Design Engineering, Imperial College London, SW7 2AZ London, U.K. CORRESPONDING AUTHOR: DAVIDE FANTINI (email: [davide.fantini@unimi.it](davide.fantini@unimi.it)). This work was supported by SONICOM, through European Union’s Horizon 2020 Research and Innovation Programme under Grant 101017743 . This article has supplementary downloadable material available at [https://doi.org/10.1109/OJSP.2025.3528330](https://doi.org/10.1109/OJSP.2025.3528330), provided by the authors. ABSTRACT Machine learning (ML) has become pervasive in various research ﬁelds, including binaural synthesis personalization, which is crucial for sound in immersive virtual environments. Researchers have mainly addressed this topic by estimating the individual head-related transfer function (HRTF) . HRTFs are utilized to render audio signals at speciﬁc spatial positions, thereby simulating real-world sound wave interactions with the human body. As such, an HRTF that is compliant with individual characteristics enhances the realism of the binaural simulation. This survey systematically examines the HRTF individualization works based on ML proposed in the literature. The analyzed works are organized according to the processing steps involved in the ML workﬂow, including the employed dataset, input and output types, data preprocessing operations, ML models, and model evaluation. In addition to categorizing the works of the existing literature, this survey discusses their achievements, identiﬁes their limitations, and outlines aspects that require further investigation at the crossroads of research communities in acoustics, audio signal processing, and machine learning. INDEX TERMS HRTF individualization, machine learning, spatial audio, binaural synthesis. I. INTRODUCTION Machine learning (ML) can be deﬁned as the learning of algorithms to solve a speciﬁc problem based on information extracted from previous experiences or events, rather than explicitly programming the algorithm [1] . ML has become pervasive in several aspects of society over the past few years, with both industrial and scientiﬁc applications. The ﬁeld of spatial audio is no exception. Spatial audio techniques ﬁnd several applications, including video gaming, teleconferencing, art, ﬂight simulation [2], devices for blind people [3], and audio production [4] . An appropriate spatial audio simulation involves the simulation of the spatial cues used by humans to localize sound sources in space. These spatial cues originate from the interactions between the human body and the sound waves, which result in position-dependent sound alterations. Head-related transfer functions (HRTFs) describe these spatial cues as a linear time-invariant (LTI) system for each sound source position of interest and for each ear. The use of an HRTF of a speciﬁc position to spatialize an audio signal spatialized with an HRTF of a speciﬁc position through headphones artiﬁcially creates the sensation of a sound source in that position. HRTFs are individual due to their close relationship with anatomical traits. Therefore, the use of an HRTF non-compliant with the individual anatomy, i.e., a non-individual HRTF, results in an improper spatial audio experience [5], [6], [7], [8], [9], [10], [11], [12] . Nonindividual HRTF

FAQs

What problem does the survey address?

The survey addresses ML-based head-related transfer function (HRTF) individualization, aiming to estimate individual-compliant HRTFs for improved binaural and spatial audio realism.

How does the survey organize the reviewed works?

The reviewed ML works are organized according to the ML workflow processing steps, including datasets, input/output types, preprocessing, model architectures, and model evaluation.

Why is standardization difficult in HRTF individualization research?

The document notes that methods are often proposed with no common validation procedure, lacking consistent datasets, objective metrics, and rigorous perceptual tests, which makes comparisons difficult.

Aug 3, 2026

Upload by Seraphina