Optimizing Large Language Models with Automatic Speech Recognition for Medication Corpus in Low-Resource Healthcare Settings.

15 Sept 2024 (modified: 05 Feb 2025)Submitted to ICLR 2025EveryoneRevisionsBibTeXCC BY 4.0
Keywords: Automatic Speech Recognition, Large Language Models, Healthcare, Low Resource Settings
Abstract:

Automatic Speech Recognition (ASR) systems, while effective in general contexts, often face challenges in low-resource settings, especially in specialized domains such as healthcare. This study investigates the integration of Large Language Models (LLMs) with ASR systems to improve transcription accuracy in such environments. Focusing on medication-related conversations in healthcare, we fine-tuned the Whisper-Large ASR model on a custom dataset, Pharma-Speak, and applied the LLaMA 3 model for second-pass rescoring to correct ASR output errors. To achieve efficient fine-tuning without altering the full LLM parameters, we employed Low-Rank Adaptation (LoRA), which enables re-ranking of the ASR’s N-best hypotheses while retaining the LLM's original knowledge. Our results demonstrate a significant reduction in Word Error Rate (WER) across multiple epochs, validating the effectiveness of the LLM-based rescoring method. The integration of LLMs in this framework shows potential for overcoming the limitations posed by conventional ASR models in low-resource settings. While computational constraints and the inherent strength of Whisper-Large presented some limitations, our approach lays the groundwork for further exploration of domain-specific ASR enhancements using LLMs, particularly in healthcare applications.

Primary Area: generative models
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 926
Loading

OpenReview is a long-term project to advance science through improved peer review with legal nonprofit status. We gratefully acknowledge the support of the OpenReview Sponsors. © 2025 OpenReview