<![CDATA[Accurate classification of singers into Soprano, Alto, Tenor, and Bass (SATB) voice categories is essential for choir performance and vocal health. However, existing automatic SATB classification methods are constrained by the limited availability of labeled datasets. This study proposes a statistically guided pseudo-labeling framework to expand labeled singing voice data and improve automatic SATB classification. Fundamental frequency (F0) was extracted from professionally labeled recordings in the Choral Singing Dataset (CSD) and ESMUC Choir Dataset using the YIN algorithm. Ten F0-derived statistical features were evaluated through statistical significance testing, effect size analysis, overlap analysis, and single-feature classification to identify the most discriminative feature for threshold-based pseudo-label generation. The selected feature was then used to assign SATB pseudo-labels to 3,093 VocalSet recordings, producing a combined dataset of 3,188 labeled recordings. The expanded dataset was segmented into four-second audio chunks, converted into Mel-spectrograms, and used to train and compare Convolutional Neural Network (CNN) and CNN–Long Short-Term Memory (CNN-LSTM) models. Experimental results identified the 75th percentile (P75) of F0 as the most discriminative statistic, achieving 98.95% single-feature classification accuracy. On the internal test set, the CNN-LSTM model outperformed the baseline CNN, achieving 94.6% accuracy and a macro F1-score of 0.92, compared with 92.6% accuracy and a macro F1-score of 0.90 for CNN. External evaluation on previously unseen YouTube recordings further demonstrated superior generalization, with CNN-LSTM achieving 65.4% accuracy versus 42.5% for CNN. These findings demonstrate that statistically guided pseudo-labeling effectively expands labeled training data and that integrating temporal modeling significantly improves automatic SATB singing voice classification.]]>
Copyrights © 2026
