Accurately identifying the working state of railway switch machines can prevent the failures from affecting the efficiency and safety of railway transportation. We proposed a state recognition method to estimate the working state for switch machines based on acoustic analysis by improving the MFCC method and obtained multi-scale MFCC (Ms-MFCC) features, which provided a more comprehensive understanding of the switch-machine acoustic signals. Convolutional neural network (CNN) was introduced to construct the acoustic signal recognition model, and the five-fold cross validation method was employed to obtain the recognition accuracy. In this study, the real acoustic signals collected by the S700K switch machine in four states were trained and tested. Our results show that the Ms-MFCC feature can improve the accuracy of the switch machine's acoustic state recognition by at least 7.5%. And in low SNR cases, the accuracy can be improved by 35%.