Domain Specific Intent Classification with BiLSTM

In this work, we propose an Automatic Speech Recognition System (ASR) with Bi-directional Long-Short-Term Memory Network (BiLSTM) that uses transfer learning and end-to-end deep learning for the healthcare domain. The domain-specific classification solutions have gained more attention in recent years as the research of the open domain is really costly as creating a good quality dataset comes with a high cost. This is one of the main challenges faced by low-resource languages like Sinhala or domains like healthcare due to privacy issues. Most of the existing research has used a text-based intermediate form of the spoken clip and performs the classification on top of the converted text. However, the proposed solution uses end-to-end deep learning and transfer-learning as fine-tuning to improve the performance alone with data augmentation. We achieved the best performance on Sinhala banking dataset [1] with 98.53% accuracy and the performance with only using half of the epochs, unlike the benchmark. When applied to the publicly available dataset AudioMNist [21] has recorded 98.54% of accuracy over 5-fold cross-validation.