Researchers developed a tone-conditioned curriculum framework for six Southern Bantu languages to address the high zero-shot word error rates (WER) of current foundation ASR models. The approach combines hybrid difficulty scoring, gated adapters driven by tonal statistics, and staged curriculum training on a community corpus.
- W2V-BERT with tone conditioning achieved an average WER of 28.41% across datasets and 23.79% on Xitsonga transfer.
- W2V-BERT outperformed Whisper by 3 to 4 WER points on Nguni languages, while Whisper performed better on Sotho-Tswana languages.
- No single model suited all six languages, indicating that deployment should pair model selection per language with validation across corpora.
The study highlights clear interactions between architecture and language, suggesting that effective deployment requires selecting specific models for each language rather than relying on a universal solution.