Listener-Variance Modeling for Calibrating Synthetic Singing Quality Scores
International Journal of Information and Communication Technology Trends · 2026
DOI: 10.66525/ijictt131 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Information and Communication Technology Trends, 2026. Verfasst von Lars D. Dijkstra und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
The evaluation of synthetic singing voices heavily relies on subjective human assessments, typically aggregated through the Mean Opinion Score paradigm. However, the perceptual evaluation of singing is notoriously complex, involving multi-dimensional cognitive judgments concerning pitch accuracy, timbral naturalness, musicality, and lyrical intelligibility. This complexity exacerbates the inherent variability among human evaluators, leading to listener-variance phenomena that obscure the true acoustic performance of underlying synthesis systems. Evaluators exhibit diverse baseline strictness, scale utilization, and internal consistency, all of which introduce substantial noise into the evaluation pipeline. To address these systemic inadequacies, this paper proposes a comprehensive listener-variance modeling framework designed to calibrate subjective quality scores in synthetic singing assessment. By conceptualizing the evaluation process through a latent variable paradigm, the proposed methodology disentangles the true systemic quality of the synthesized audio from the subjective biases and unreliability of individual listeners. The study conducts extensive perceptual experiments involving multiple generative architectures and a demographically diverse listener pool. Through sophisticated statistical calibration techniques, we demonstrate that modeling annotator behavior significantly reduces confidence intervals, alters systemic rankings to better reflect true acoustic fidelity, and enhances the statistical power of subjective evaluations. The findings underscore the critical necessity of transitioning from naive arithmetic averaging toward probabilistically calibrated scoring mechanisms in the domain of musical audio synthesis.
Zitieren
APA
Dijkstra, L. D., & Prins, H. (2026). Listener-Variance Modeling for Calibrating Synthetic Singing Quality Scores. International Journal of Information and Communication Technology Trends. https://doi.org/10.66525/ijictt131
Vancouver
Dijkstra LD, Prins H. Listener-Variance Modeling for Calibrating Synthetic Singing Quality Scores. International Journal of Information and Communication Technology Trends. 2026. doi:10.66525/ijictt131
BibTeX
@article{Dijkstra2026,
author = {Dijkstra, Lars D. and Prins, Hannah},
title = {Listener-Variance Modeling for Calibrating Synthetic Singing Quality Scores},
journal = {International Journal of Information and Communication Technology Trends},
year = {2026},
doi = {10.66525/ijictt131}
}
EUVIMED