A Comparative Study on using English ASR Models for Leveraging Medical Speech Recognition System

Though many deep learning based Automatic Speech Recognition algorithms (ASR) exist, developing a Medical Speech Recognition (MSR) system has many challenges. One such difficulty is that it requires huge annotated data for training a machine learning model to understand the medical vocabulary. As these corpora are not available publicly, this poses a great challenge in MSR development. Recently transfer learning has become popular where a pre-trained model for a task can be leveraged to do another similar task. This work examines the use of pretrained English ASR models to develop a MSR model. A controlled study has been carried out to find the potential opportunities and gaps in using English ASR models such as wav2vec, XLSR and Deepspeech in identification of medical terms from medical sentences. As these models work well for general English transcription, these models are tested with medical sentences spoken by a single speaker with varying tones and speeds and the word error for medical terms is evaluated. Among the models tested, XLSR a multilingual model outperforms the others in identifying the medical terms. From this study it is observed that building a MSR with limited data is feasible by fine-tuning specific layers with transfer learning approach.