The operating sound signal of the power transformer contains rich equipment status information, and the voiceprint signal has the characteristics of portable collection and easy analysis. However, most of the current transformer fault diagnosis methods based on acoustic signals rely on the traditional fixed acoustic characteristic frequency, and the models used also have a huge number of parameters. Therefore, this paper proposes a lightweight fault diagnosis model based on multi-voiceprint feature fusion Vision Transformer (VMF-ViT). Firstly, the frequency domain gram angular field map and the chaotic feature map are constructed. Then, the ViT is used to extract transformer state features. Finally, the fault mode recognition results are output by using the classifier. The experimental verification results show that the proposed method with a diagnosis accuracy of more than 97%, and the model parameters and calculation are less and the fault diagnosis accuracy is higher than that of the typical neural network diagnosis algorithm and transformer voiceprint fault diagnosis algorithm.