Researchers analyzed how multilingual medical adaptation affects the internal representations of a popular automatic speech recognition (ASR) model called Whisper. They found that fine-tuning the model improves performance in medical ASR tasks, but the best approach depends on the specific adaptation setting and language. The study used various methods to adapt the model, including zero-shot decoding, English-only fine-tuning, and direct multilingual training. Layer-wise analysis showed that English medical fine-tuning has a significant impact on the model's internal representations, while multilingual continuation preserves the adapted representation space. --- Why it matters: This research matters because it sheds light on how to adapt ASR models for medical applications in multiple languages, which is crucial for improving healthcare services globally. Engineers and researchers can use these findings to develop more effective methods for fine-tuning ASR models for specific languages and domains. Source: https://arxiv.org/abs/2608.18825