Speech Error Detection depending on Linguistic Units

Abstract

In this research, we aim at the construction of a system which detects, points out and corrects speech error (slip of the tongue) of a human speech that occurs in a dialogue system (example: Pepper, Amazon Echo, Google Home) and a human dialogue. In the present dialogue system, even if human makes a speech error, the system cannot recognize it, which could lead to broken communication. So far, we have created a system to detect speech error using deep learning. In this study, we propose a method to augmented training data used for deep learning. The training data is a corpus that collects examples of speech error. At present, the number of training data is insufficient to detect with high accuracy. Therefore, it is necessary to augment the training data. Specifically, the feature of the speech error is examined from an existing speech error corpus, and extended rules are created. The data augmentation of training data is performed by generating dialogue sentence which made the speech error based on the rule. As a result of evaluation experiment, detection accuracy was improved in LSTM model by data augmentation.