Neural audio synthesis technologies have rapidly become ubiquitous in the music and multimedia industries, causing widespread sociocultural, economic, and aesthetic effects. The evaluation stage is a particularly relevant phase of their development, as it defines the implicit aesthetic goals a model should pursue. This study examines sixty-eight recently designed neural audio synthesis models, critically inquiring into their objective and subjective evaluation criteria. Reductionist assumptions, incomplete information, limited experimental replicability, and small sample sizes significantly compromise the epistemic reliability of current evaluation practices. Similarity to pre-existing music and vague notions of quality and musicality emerge as underlying, poorly articulated aesthetic principles. Neural audio engineering inherits its evaluation methodologies from pre-existing, non-creative research fields, uncritically applying them to aesthetic tasks such as automatic music generation. Consequently, the role of listening is merely reduced to numerical grading, rejecting any proper qualitative assessment and flattening the aesthetic experience into uni-dimensional quantitative metrics. In order to overcome these epistemic inadequacies, engineers and musicologists should cooperate to complement thorough objective measurements with phenomenologically-grounded forms of listening.
Ancona, R. (2026). Epistemic Inadequacies in Neural Audio Synthesis Evaluation [10.5281/zenodo.22255971].
Epistemic Inadequacies in Neural Audio Synthesis Evaluation
Ancona R.
2026
Abstract
Neural audio synthesis technologies have rapidly become ubiquitous in the music and multimedia industries, causing widespread sociocultural, economic, and aesthetic effects. The evaluation stage is a particularly relevant phase of their development, as it defines the implicit aesthetic goals a model should pursue. This study examines sixty-eight recently designed neural audio synthesis models, critically inquiring into their objective and subjective evaluation criteria. Reductionist assumptions, incomplete information, limited experimental replicability, and small sample sizes significantly compromise the epistemic reliability of current evaluation practices. Similarity to pre-existing music and vague notions of quality and musicality emerge as underlying, poorly articulated aesthetic principles. Neural audio engineering inherits its evaluation methodologies from pre-existing, non-creative research fields, uncritically applying them to aesthetic tasks such as automatic music generation. Consequently, the role of listening is merely reduced to numerical grading, rejecting any proper qualitative assessment and flattening the aesthetic experience into uni-dimensional quantitative metrics. In order to overcome these epistemic inadequacies, engineers and musicologists should cooperate to complement thorough objective measurements with phenomenologically-grounded forms of listening.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.



