A Sequence-to-sequence Based Error Correction Model for Medical Automatic Speech Recognition
Yu Lian Jiang, Christian Poellabauer · 2021 IEEE International Conference on Bioinformatics and Biomedicine (BIBM) · 2021
The use of Automatic Speech Recognition (ASR) systems in medical applications is receiving rapidly growing interest due to their ability to reduce distractions and the cognitive workload of physicians, particularly during critical medical procedures. However, state-of-the-art ASR systems still experience recognition errors, especially in noisy environments where speakers rely on medical-domain terminologies. This paper proposes a customized language model and a neural network based sequence-to-sequence (seq2seq) error correction module for medical ASR systems to provide domain adaptation and more reliable transcription results. Specifically, the error correction module learns the error patterns in noisy scenarios and is able to correct such errors during inference. Our experiments show that the proposed method can reduce the sentence error rate (SER) by up to 81% for formatted input and up to 31% SER for unformatted input in noisy environments.