Automatic Speech Recognition Model Adaptation to Medical Domain Using Untranscribed Audio

Automatic speech recognition (ASR) technologies can provide significant efficiency gains in the health sector, by saving time and financial resources, allowing specialists to shift more time to high-value activities. Creating customized ASR models requires domain- and task-related transcribed speech data. Unfortunately, producing such data usually is too expensive for medical institutions: it requires a lot of financial, human resources, and expertise. Consequently, his paper explores a semi-supervised medical domain adaptation method for the Latvian language that benefits from the untranscribed speech recordings. For the initial model, we use the currently available general-purpose hybrid ASR system with the core of a lattice-free maximum mutual information method used to train its acoustic model. The initial system is applied to the domain-related untranscribed data to extract sequences of pseudo-labels. Such automatic transcriptions are later added to the supervised and used together to update the acoustic model. To improve our ASR system further, we have also updated its language model with additional in-domain texts. We have achieved significant improvements in the quality of speech recognition on all evaluation datasets. On the epicrises, psychiatry, and radiology datasets word error rate (WER) decreased by 39%, 27%–29%, and 21%, respectively.