2021/08/13 by Ewald van der Westhuizen, van der Westhuizen, Ewald, Trideba Padhi +3 · 1 citation
Computer Science · #Audio and Speech Processing (eess.AS) #FOS: Electrical engineering #Natural Language Processing Techniques #Speech Recognition and Synthesis #Speech and dialogue systems #electronic engineering #information engineering
paper · pdf · doi:10.48550/arxiv.2108.06164
openalex publication_date 2021/08/13 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
We present first speech recognition systems for the two severely\nunder-resourced Malian languages Bambara and Maasina Fulfulde. These systems\nwill be used by the United Nations as part of a monitoring system to inform and\nsupport humanitarian programmes in rural Africa. We have compiled datasets in\nBambara and Maasina Fulfulde, but since these are very small, we take advantage\nof six similarly under-resourced datasets in other languages for multilingual\ntraining. We focus specifically on the best composition of the multilingual\npool of speech data for multilingual training. We find that, although\nmaximising the training pool by including all six additional languages provides\nimproved speech recognition in both target languages, substantially better\nperformance can be achieved by a more judicious choice. Our experiments show\nthat the addition of just one language provides best performance. For Bambara,\nthis additional language is Maasina Fulfulde, and its introduction leads to a\nrelative word error rate reduction of 6.7%, as opposed to a 2.4% relative\nreduction achieved when pooling all six additional languages. For the case of\nMaasina Fulfulde, best performance was achieved when adding only Luganda,\nleading to a relative word error rate improvement of 9.4% as opposed to a 3.9%\nrelative improvement when pooling all six languages. We conclude that careful\nselection of the out-of-language data is worthwhile for multilingual training\neven in highly under-resourced settings, and that the general assumption that\nmore data is better does not always hold.\n