A Dataset for Automatic Vocal Mode Classification

TL;DR

The paper addresses the challenge of automatic vocal mode classification in the Complete Vocal Technique (CVT) by creating a publicly available, multi-singer dataset covering all four CVT modes across the full vocal range. It combines sustained vowel recordings from four singers using four recording devices, with ground-truth annotations from three CVT-informed annotators and a merged annotation, enabling robust evaluation via $5$-fold cross-validation and balanced accuracy. Baseline results show ResNet18 achieving $81.3\%$ balanced accuracy with the full annotation, while using nominal ground truth yields $95.3\%$ with ResNet34, and XGBoost reaches $92.8\%$ in the nominal setting, highlighting strong gains when ground truth aligns with perceptual labels. The dataset, available at Zenodo, provides a valuable resource for CVT-related research and technology-assisted singing pedagogy, while the authors discuss annotation quality and limitations, pointing to future work with more singers and realistic musical material.

Abstract

The Complete Vocal Technique (CVT) is a school of singing developed in the past decades by Cathrin Sadolin et al.. CVT groups the use of the voice into so called vocal modes, namely Neutral, Curbing, Overdrive and Edge. Knowledge of the desired vocal mode can be helpful for singing students. Automatic classification of vocal modes can thus be important for technology-assisted singing teaching. Previously, automatic classification of vocal modes has been attempted without major success, potentially due to a lack of data. Therefore, we recorded a novel vocal mode dataset consisting of sustained vowels recorded from four singers, three of which professional singers with more than five years of CVT-experience. The dataset covers the entire vocal range of the subjects, totaling 3,752 unique samples. By using four microphones, thereby offering a natural data augmentation, the dataset consists of more than 13,000 samples combined. An annotation was created using three CVT-experienced annotators, each providing an individual annotation. The merged annotation as well as the three individual annotations come with the published dataset. Additionally, we provide some baseline classification results. The best balanced accuracy across a 5-fold cross validation of 81.3\,\% was achieved with a ResNet18. The dataset can be downloaded under https://zenodo.org/records/14276415.

Figures (5)

• Figure 1: Number of samples per note for the entire dataset. The highest note, a C6, was achieved in nominal Neutral. The lowest note, an F1, was achieved also in nominal Neutral.
• Figure 2: Number of samples per subject and empirical cumulative distribution of the sample durations in seconds of all files of the dataset.
• Figure 3: Balanced accuracies on the test set across the 5-fold cross validation of all investigated classifiers using (a) the annotated vocal mode and (b) the nominal vocal mode as ground truth.
• Figure 4: Balanced accuracy across half-octaves on the test set for the best iterations of the 5-fold cross validation when (a) using the full annotation and (b) using the nominal vocal modes as ground truth. The very high accuracy for the top end of the range in (b) is largely due to a lack of data. Correspondingly, the drop to 50 % balanced accuracy in (a) is due to a lack of data in this particular iteration, where only for two vocal modes were present for the highest range. Qualitative identical behavior was observed for all classifiers.
• Figure 5: Fleiss' kappa score across cut-off note threshold. The computation of the Fleiss' kappa score given a note threshold of K ignores the lowest K notes of each singer, irrespective of the vocal mode.