Beyond the Feature: Comparing Spectral, Cepstral, and Self-Supervised Representations for Sesotho Tone Classification, and What It Means for a Southern African Methodological Commons
Journal of the Digital Humanities Association of Southern Africa (DHASA) · 20.11.2026 · 7 (2) · 67-78
DOI: 10.55492/dhasa.v7i02.7696 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of the Digital Humanities Association of Southern Africa (DHASA), online erschienen am 20.11.2026. Verfasst von Tshepo Mashishi und 2 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY-SA 4.0).
Abstract (Originalsprache)
Sesotho, a Southern Bantu language spoken by roughly 5.6 million people in South Africa and Lesotho, uses a binary High/Low pitch system to distinguish word meanings and grammatical forms (for example, the person of a verb’s subject), a system that most speech and language technologies are not built to recognise. This paper reports a controlled comparison of four feature-extraction paradigms for automated Sesotho tone classification – Fast Fourier Transform (FFT) spectral features, a Librosa acoustic pipeline, Mel-Frequency Cepstral Coefficients (MFCC), and Wav2Vec2 self-supervised embeddings – each evaluated with five classifiers on 4,056 syllable segments from 15 speakers, of which 1,278 carry expert tone labels. Evaluation uses a speaker-disjoint split: 11 speakers for training and four held-out speakers (297 syllables) for testing. A 156-dimensional MFCC feature set paired with a neural network achieves the best held-out result (F1 = 0.9190, ROC-AUC = 0.9583), ahead of a 24-dimensional FFT feature set (F1 = 0.8718, ROC-AUC = 0.9358), Wav2Vec2 (F1 = 0.7966, ROC-AUC = 0.8938), and the Librosa baseline (F1 = 0.6916, ROC-AUC = 0.7740). The ranking holds when classifiers are chosen by cross-validation instead of on the test set. Feature dimensionality alone does not predict performance: the 24-dimensional FFT set outperforms the 776-dimensional Wav2Vec2 set and trails MFCC by 0.047 F1. The feature representation matters more than the classifier: the gap between the best and worst feature sets (0.227 F1) exceeds the spread across classifiers within any feature set (0.041 − −0.072). An earlier MFCC extraction lacked features for most syllables; the results reported here use a corrected extraction. Beyond these findings, we situate the work within the Digital Humanities Association of Southern Africa’s call for a methodological commons, arguing that reproducible, low-resource benchmarking of this kind is a form of infrastructure-building for linguistic justice in the region.
Zitieren
APA
Mashishi, T., Senekane, M., & Primus, M. (2026). Beyond the Feature: Comparing Spectral, Cepstral, and Self-Supervised Representations for Sesotho Tone Classification, and What It Means for a Southern African Methodological Commons. Journal of the Digital Humanities Association of Southern Africa (DHASA), 7(2), 67-78. https://doi.org/10.55492/dhasa.v7i02.7696
Vancouver
Mashishi T, Senekane M, Primus M. Beyond the Feature: Comparing Spectral, Cepstral, and Self-Supervised Representations for Sesotho Tone Classification, and What It Means for a Southern African Methodological Commons. Journal of the Digital Humanities Association of Southern Africa (DHASA). 2026;7(2):67-78. doi:10.55492/dhasa.v7i02.7696
BibTeX
@article{Mashishi2026,
author = {Mashishi, Tshepo and Senekane, Makhamisa and Primus, Mpho},
title = {Beyond the Feature: Comparing Spectral, Cepstral, and Self-Supervised Representations for Sesotho Tone Classification, and What It Means for a Southern African Methodological Commons},
journal = {Journal of the Digital Humanities Association of Southern Africa (DHASA)},
year = {2026},
volume = {7},
number = {2},
pages = {67--78},
doi = {10.55492/dhasa.v7i02.7696}
}
EUVIMED