The NTNU System at the S&I Challenge 2025 ASR Open Track

This paper explores an effective approach to advancing nonnative (L2) automatic speech recognition (ASR). To this end, we design and implement a pragmatic architecture for L2 ASR and use an effective fine-tuning mechanism tailored for L2 ASR. Experiments conducted on the Speak & Improve 2025 corpus demonstrate that our approach can offer a significant improvement of 53.7% in terms of SpWER compared to the baseline. Building on top of the proposed approach, we came out in first place on the open track of the ASR task in Speak & Improve Challenge 2025 (S&I Challenge).