A computer-implemented method for utterance generation, a smart device, and a non-transitory computer readable storage medium are provided. The method includes: obtaining a first utterance to be answered, generating at least one random semantic vector, inputting the at least one random semantic vector and the first utterance into a trained generator, and obtaining at least one first answer outputted by the trained generator, wherein the trained generator is obtained based on a preset generative adversarial network. Due to the random semantic vector, even for the same utterance, the smart device can generate different answers corresponding to the different random semantic vectors, the possibility of generating too many identical answers during the human-machine conversation is reduced, and the fun during the human-machine conversation is enhanced.