Distorted Realities: Classifying Extreme Vocals Between Harmony and Noise – A Machine–Human Evaluation of Vocal Confusion Patterns

Extreme vocal techniques like growling and shrieking are common in metal and related genres, giving the music a characteristic intensity. Although widely used, these techniques have received limited attention in Music Information Retrieval (MIR) and musicology, particularly in fine‑grained categorisation. Their niche status and marginalisation have led to limited representation in digital music libraries. Since current classification systems prioritise dominant subgenres due to a lack of detailed vocal categorisations, extreme vocal styles are often oversimplified or misrepresented, hindering tagging, retrieval, and analysis. Automation could enable consistent and nuanced vocal annotation, supporting the technical goals of MIR and analytical aims of musicology. Foregrounding vocal technique as a classification axis contributes to a richer understanding of vocal performance and genre formation. To assess machine learning as a reliable alternative to human judgment, this study developed a coarse frequency‑based qualitative taxonomy from interviews with 11 extreme vocal practitioners and compared human and machine performance using the Extreme Metal Vocals Dataset (EMVD), including a listening test with 158 expert participants. Perceptual results achieved 76.2% Unweighted Average Recall (UAR), while a Support Vector Machine (SVM) model trained on ComParE features achieved a UAR of 90% on average, using singer‑independent 3‑fold group cross‑validation. Feature group, ablation, and human prediction analyses suggest that the human‑machine gap reflects differences in cue prioritisation and classification performance. Results illustrate the feasibility of machine learning to annotate vocal styles, while showing that human confusion patterns remain informative for category boundaries. This highlights how MIR methods could facilitate analysis of underrepresented genres.