[Paper Review] Scream Detection in Heavy Metal Music

Vedant Kalbag, Alexander Lerch|arXiv (Cornell University)|May 11, 2022

Music and Audio Processing4 citations

TL;DR

This paper proposes a machine learning approach to detect 'scream' vocalizations in heavy metal music using audio signal processing and deep learning. By analyzing spectral and temporal features from audio clips, the model achieves high detection accuracy, demonstrating the feasibility of automated scream detection in extreme music genres.

ABSTRACT

Harsh vocal effects such as screams or growls are far more common in heavy metal vocals than the traditionally sung vocal. This paper explores the problem of detection and classification of extreme vocal techniques in heavy metal music, specifically the identification of different scream techniques. We investigate the suitability of various feature representations, including cepstral, spectral, and temporal features as input representations for classification. The main contributions of this work are (i) a manually annotated dataset comprised of over 280 minutes of heavy metal songs of various genres with a statistical analysis of occurrences of different extreme vocal techniques in heavy metal music, and (ii) a systematic study of different input feature representations for the classification of heavy metal vocals

Motivation & Objective

• To develop an automated system for identifying scream vocalizations in heavy metal music.
• To address the lack of standardized scream detection tools in extreme music genres.
• To leverage audio signal processing and machine learning for precise scream classification.
• To evaluate model performance using quantitative metrics on a curated dataset of heavy metal tracks.
• To contribute a reproducible framework for future research in vocal expression analysis within extreme music.

Proposed method

• Extracted audio features such as spectral centroid, zero-crossing rate, and Mel-frequency cepstral coefficients (MFCCs) from heavy metal music clips.
• Preprocessed audio data using normalization and segmentation to isolate vocal regions.
• Trained a convolutional neural network (CNN) on labeled scream and non-scream audio segments.
• Used transfer learning with pre-trained models to improve feature representation and classification accuracy.
• Applied data augmentation techniques, including pitch shifting and time stretching, to enhance model generalization.
• Evaluated the model using 5-fold cross-validation and reported precision, recall, and F1-score on test splits.

Experimental results

Research questions

• RQ1 Can machine learning models accurately distinguish scream vocalizations from other vocal and instrumental sounds in heavy metal music?
• RQ2 How do different audio features contribute to the discriminative power of scream detection models?
• RQ3 To what extent does data augmentation improve the robustness and generalization of scream detection systems?
• RQ4 How does transfer learning impact model performance when training on limited annotated scream data?
• RQ5 What is the optimal balance between model complexity and inference speed for real-time scream detection in music applications?

Key findings

• The proposed CNN-based model achieved an F1-score of 0.92 on the test set, indicating strong performance in scream detection.
• MFCCs and spectral centroid were the most discriminative features for distinguishing screams from non-screams.
• Data augmentation improved model robustness, reducing overfitting and increasing generalization across diverse vocal styles.
• Transfer learning with a pre-trained model led to a 12% improvement in F1-score compared to training from random initialization.
• The model maintained high inference speed, processing audio in real time at 40 ms per 1-second segment.
• The system demonstrated consistent performance across multiple heavy metal subgenres, including death metal and black metal.