Tone-Conditioned Curriculum Learning for Low-Resource Bantu Speech Recognition
Researchers developed a framework to improve speech recognition for Southern Bantu languages by combining Word Error Rate (WER) with tonal features into a hybrid difficulty scoring function. They used tone-conditioned gated adapters during staged fine-tuning. The study found that architecture suitability varies by language family, with W2V-BERT performing best on Nguni languages, and tone conditioning helped reduce errors for some models.