In the field of speech recognition, performance varies much when the system is trained or tested with different data. In this paper, we explore the effect of training and test data on the performance of automatic speech recognition systems. Unlike other researchers who analyze the effect of training and testing as pattern learning and recognition of vectors, the effect of data is investigated as effect of data properties, such as SNR and kind of environmental noise. For a data property, a statistical model based on ANOVA was proposed to decompose the effect on system performance into three parts - effect of training data, test data and their interaction, and each part is considered dependent on the level of data properties. Experiments were conducted on a LVCSR system for the data properties of kind of noise and SNR, and results and analysis are presented to explain how they influence the performance by training and test.