This paper aims to investigate the dynamic splitting tensile damage characteristics of steel fiber reinforced concrete (SFRC) and achieve early warning of the dangerous stage under freeze-thaw cycle environment through acoustic emission (AE) technology, Mel-frequency cepstral coefficients (MFCCs) and genetic algorithm-back propagation (GA-BP) neural network. The effects of fiber content and loading rate on mechanical properties, wavelet energy spectrum coefficient and wavelet time-frequency-energy maximum, as well as MFCCs of AE signals were studied herein. Then a GA-BP neural network was established to realize the classification of the damage degree of SFRC under freeze-thaw cycling with MFCCs as the signal characteristic. The results indicate that the increasing loading rate leads to more serious brittle failure, but the appropriate steel fiber content can make the cracks generated develop fully during the dynamic splitting tensile failure process by analyzing the wavelet energy spectrum coefficient and wavelet time-frequency-energy maximum. It is also found that the damage degree of SFRC can be identified by analyzing the MFCCs. Furthermore, the GA-BP neural network performs well in the classification of the damage degree of SFRC with the recognition over 90%, which is more suitable for concrete conditions with a proper loading rate and fiber content.