Unifying Heterogeneous HRTF Databases: A Data-Driven and Geometry-Aware Approach to Explainable Personalized Spatial Audio

• Autores: Juan Antonio de Rus Arance
• Directores de la Tesis: Mario Montagud Climent (dir. tes.), Máximo Cobos Serrano (codir. tes.)
• Lectura: En la Universitat de València ( España ) en 2026
• Idioma: inglés
• Tribunal Calificador de la Tesis: Miguel Lozano Ibáñez (presid.), José Javier Valero Mas (secret.), Federica Battisti (voc.)
• Programa de doctorado: Programa de Doctorado en Tecnologías de la Información, Comunicaciones y Computación por la Universitat de València (Estudi General)
• Menciones:
  • Internacional
• Calificación: Sobresaliente
• Texto completo no disponible (Saber más ...)
• Resumen

  • This thesis addresses spatial-audio personalization in immersive environments, with a particular focus on binaural reproduction based on Head-Related Transfer Functions (HRTFs). Although individualized HRTFs are essential for realistic spatial perception, their direct measurement remains costly and difficult to scale. The thesis proposes a comprehensive, interpretable, and deployment-oriented framework connecting the analysis of elevation-related spectral cues, the geometric representation of pinna anatomy, personalized HRTF synthesis, and their evaluation in immersive applications.

    The first part investigates elevation cues in HRTFs through machine learning and explainable artificial intelligence. A harmonized multi-dataset framework is established to address the heterogeneity of public HRTF databases regarding angular conventions, preprocessing, and acquisition conditions. Convolutional models are trained for elevation classification, and their decisions are examined through saliency analysis. Results show that training diversity improves cross-dataset robustness and that the most informative regions are mainly concentrated in high-frequency bands, although relevant lower-frequency contributions also appear depending on source direction. The thesis further validates these salient regions through controlled perceptual experiments based on spectral ablation, showing that their manipulation causes measurable degradations in vertical sound localization.

    The second part focuses on geometry-based personalization. A unified representation of 3D pinna geometry is proposed to resolve the heterogeneity of available ear meshes, including differences in topology, density, scope, and vertex ordering. The pipeline combines preprocessing, anatomically informed landmarking, non-rigid registration, and topology-preserving reindexing, yielding a compact and correspondence-consistent geometric descriptor across subjects. Building on this representation, the thesis studies HRTF synthesis from geometry by comparing signal domains, compression strategies, and loss formulations. Results indicate that standardized geometry simplifies learning and improves spectral fidelity over approaches based on unregistered meshes. Compact low-dimensional descriptors retain sufficient anatomical information to achieve competitive results, including under cross-dataset generalization.

    The thesis is complemented by appendices oriented toward deployment and evaluation in realistic usage scenarios. It develops software tools for reproducible geometry processing, SOFA inspection and comparison, and real-time monitoring of spatial-audio pipelines. It also implements an immersive Unity/VR environment capable of loading measured, generic, or synthesized HRTFs and evaluating them through objective localization tasks and gamified Quality of Experience (QoE) scenarios. The main contribution of this block lies in the experimental and methodological infrastructure already in place, rather than in a completed large-scale comparative campaign. Finally, the thesis explores physiological sensing as a complementary layer for multimodal QoE assessment. A standardized multi-device acquisition and processing pipeline is proposed, together with a cross-device modeling strategy anchored in ECG supervision. This contribution is explicitly exploratory: it demonstrates methodological feasibility and prepares future multimodal campaigns, without claiming definitive affective validation.

    Overall, the thesis provides a unified view of spatial-audio personalization, progressing from the identification and interpretation of relevant cues, through their connection with reusable anatomical representations, to their integration into tools and immersive environments for evaluation. The result is an interpretable, geometry-aware, and deployment-oriented framework for advancing toward more robust and practically applicable personalized spatial-audio systems.