Abstract
Nowadays, many latest systems are typical cyber physical systems (CPS), such as self-driving systems, medical monitoring, industrial control systems and robotics systems. Some of these fields involve speech emotion recognition based on deep learning technology. Therefore, the safety issues brought by deep neural networks cannot be ignored. Recurrent neural network (RNN) is one of several mainstream directions in speech emotion recognition. However, limited research has been done on RNN testing. In this paper, we define important-unit coverage metric for a classic RNN architecture, long short-term memory network (LSTM), to guide the generation of test cases and measure the test adequacy. We implement our experiments on a speech emotion dataset named Emo-DB. We also compare our method with some existing test coverage metrics for RNN. Experimental results show that we have consistent performance comparing with these metrics and can generate more test cases than neuron coverage.