Abstract
Context: Automatic Speech Recognition (ASR) in healthcare is undergoing a significant shift driven by the integration of Large Language Models (LLMs). While traditional ASR focused on transcription fidelity, LLM-based systems extend this capability to intelligently reason, summarize, and structure clinical data. This scoping review maps the emerging landscape of LLM-based ASR in healthcare, examining its applications, technical foundations, evaluation practices, and reported challenges. Methods: Following PRISMA-ScR guidelines, we searched different databases for peer-reviewed, open-access studies published between January 2022 and December 2025 to ensure reproducibility and accessibility. Results: Nineteen studies met the inclusion criteria from 384 screened records. Administrative documentation was the most common application (42.1%), followed by diagnosis, therapy, and doctorβpatient communication. Whisper dominated ASR (52.6%), typically paired with GPT-family or LLaMA-family LLMs in frozen configurations steered through prompting. LLMs served as the primary component in 68.4% of studies. ASR evaluation within the reviewed studies predominantly relied on word error rate, while LLM evaluation remains fragmented with no standard metric. Studies reported documentation time reductions of 30β90%, though privacy reporting was inconsistent, equity concerns were rarely tested systematically, and only five studies provided replication packages. Conclusions: LLM-based ASR shows potential for reducing documentation burden and supporting clinical workflows, but gaps in evaluation standardization, equity testing, and reproducibility must be addressed before safe clinical deployment.