An Enhanced Model for ASR in the Medical Field

The application of speech recognition has been utilized to enhance work efficiency and even improve quality of life. Previous studies have indicated that using general-purpose speech recognition models for domain-specific recognition yields suboptimal results. Therefore, this study aims to enhance speech recognition models and optimize the correction of generated errors to improve the model’s performance in the medical field. This study employed the PubMed dataset to optimize the speech recognition language model and dictionary. Subsequently, the output results were subjected to error correction using the Bert2Bert, BioBert2BioBert, and BART model architectures. Finally, the model was tested using out-of-domain data from the MedDialog dataset to ensure its robustness. It is anticipated that in the future, the model architecture can be applied to the medical domain to enhance efficiency in scenarios such as medical consultations.