MIRG-ICAIR Conference

Permanent URI for this community

https://ir.unilag.edu.ng/handle/123456789/13010

Browse

Browsing MIRG-ICAIR Conference by Author "Ajibade A.A."

Now showing 1 - 1 of 1
  • Item
    Open Access
    WhisperMed: Fine-Tuned ASR for Enhanced Medication Communication in Clinical Settings
    (MIRG, 2024-11-25) Dere A.A.; Ajibade A.A.
    Medication management is a critical aspect of patient safety that often faces significant communication challenges, particularly in resource-constrained environments. Errors in transcribing medication information can lead to adverse drug events, which are among the most preventable causes of patient harm. Automatic Speech Recognition (ASR) systems have shown promise in mitigating these communication issues, yet they frequently struggle with domain-specific vocabularies, especially complex medical and pharmaceutical terminology. To address these challenges, we present WhisperMed, a fine-tuned version of OpenAI's Whisper Small model, specifically designed to enhance the recognition of medication-related speech in clinical settings. WhisperMed was fine-tuned using Pharma-Speak, a dataset comprising audio recordings of medication instructions and consultations, with the goal of optimizing accuracy in recognizing drug names and dosage instructions. The training involved a learning rate of 1e-05, a train batch size of 16, and mixed precision training for computational efficiency. We evaluated the model using Word Error Rate (WER), for medication names and domain-specific recognition. The model demonstrated promising performance, with a final validation loss of 0.6189 and a WER of 20.0%, showing significant improvement compared to the baseline Whisper-Small model. Our results indicate that the fine-tuned model effectively addressed medication-specific challenges, although limitations such as overfitting were noted, likely due to insufficient diversity in the training dataset and resource constraints. WhisperMed represents a step forward in enhancing medication communication in clinical environments, especially in settings with limited resources, by bridging the gap between general ASR capabilities and the specific needs of healthcare providers. This specialized ASR tool contributes towards improving the accuracy of medication transcription, ultimately aiming to reduce the risk of adverse drug events and improve patient safety.