Exploring adaptation techniques of large speech foundation models for low-resource ASR: a case study on Northern Sámi

Speech foundation models such as wav2vec 2.0 have made it possible to develop highly accurate models for low-resourced languages using a limited amount of speech data. For optimal results, the pre-training should already include data from the target language, but unfortunately, none of the available foundation models include Northern Sámi. In this work, we explore various ways of preparing the foundation model for the Northern Sámi, including continued pre-training with a small untranscribed corpus and our new extended fine-tuning method. The extended fine-tuning starts from an already fine-tuned ASR model and augments it with new output units for the unique Sámi characters before new fine-tuning with transcribed Sámi data. Our results demonstrate the benefits of these advanced adaptation techniques, as both approaches lead to better performance than the direct fine-tuning-based adaptation.