Parameter-Efficient Finetuning of ASR Models for Maritime Radio Communications in Ireland

The need for manual logging of critical Very High Frequency (VHF) radio communi-cations by the Irish Coast Guard (IRCG) puts an additional cognitive load on watch officers, compromising their situational awareness. While Automatic Speech Recognition (ASR) offers a solution, generic foundational models underperform in the maritime domain due to acoustic mismatches and a scarcity of domain-specific training data. The goal of this work is to address this gap by developing a data preparation pipeline and evaluating different parameter-efficient finetuning (PEFT) methods in a low-resource scenario. A 2-hour labelled dataset was created and used to train the OpenAI whisper-large model. The effectiveness of two PEFT techniques, namely, Encoder Freezing and Low-Rank Adaptation (LoRA) was evaluated. These finetuned models outperformed the original model’s baseline 62.8% Word Error Rate (WER). The constrained LoRA configuration proved most effective, achieving a 28.12% WER, a 52.2% relative reduction in error.