Announcement Capture System in Real Environments Using Recurrent Neural Network

Announcement is useful transmission mean. It is used in various places. Important information such as evacuation guidance in the case of emergency is often transmitted in the announcements. But some people miss it due to various factors. In this paper, we propose an announcement capture system to resolve that problems. The system consists of three steps: Firstly, announcements are detected in real environmental recordings. Secondly, duration of the announcement is extracted from detected sounds. Finally, extracted announcements are output in some form on user's device. For the first step to develop the proposed system, we validated the performance of the classifier to detect announcements. The classifier was trained in some features using BLSTM which is one of the methods of machine learning. In the validation experiment, the performances of each classifiers trained by varying features were compared. As results of the validation, the feature in the human perceptual aspect was effective to identify announcements. In addition to the result, it was considered that there is a possibility to improve the performance of announcement detection using the feature in the acoustic aspect. However, to incorporate the acoustic feature, reviewing the hyperparameters and removing surrounding sounds of the announcement are required.