Self-supervised and Multilingual Learning Applied to the Wolof, Swahili and Fongbe

Under-resourced languages face significant challenges in speech recognition due to limited resources and data availability, hampering their development and usage. In this paper, we present a speech recognition model built upon existing frameworks based on self-supervised learning (Contrastive Predictive Coding (CPC), wav2vec and bidirectional version of CPC) by combining them with multilingual learning. This model is experimented on Wolof, Swahili, and Fongbe which are African languages. The results of our evaluation of representations on the automatic speech recognition task, using a similar architecture to DeepSpeech, highlight the model’s capability to discriminate language-specific linguistic features, achieving a Word Error Rate (WER) of 61% for Fongbe, 72% for Wolof and 88% for Swahili.