Existing soundscape prediction model application conditions and calculation processes are very complex, and difficult to combine with design processes. Therefore, in this study, a visual soundscape prediction model is constructed for sound pressure level (SPL) prediction, sound source prediction, and soundscape evaluation prediction by inputting the elements of urban design under the meshing method (grid 50 m) into a machine learning Gaussian mixture model (GMM). Taking three typical urban parks as examples, the soundwalk method is used to collect subjective perception information on-site to verify prediction accuracy. The results show that by applying geographic information data, including the minimum distance values from predicted points to roads, entrances and exits, and internal nodes, the SPL can be predicted. When the accuracy rate is stable within a 3 dBA error range, the prediction accuracy rate is 67.7%. The visual perception information added is used to quickly predict the sound source type, and the output is the visualized distribution of the natural, human, and mechanical sound perception with a 77.4% accuracy rate. Finally, combining geographic, visual and sound data to predict soundscape evaluation, the model's output are good, medium, and poor categories for each descriptor of the soundscape evaluation, and the prediction results are visualized in three colours with a 74.2% accuracy rate. The convenient soundscape prediction model proposed in this paper can be applied to design practice. By adjusting park design elements, the distribution results can be compared, and an optimal design scheme from the soundscape perspective can be obtained. Combined with the model output, this study compares the model simulation results after adjusting the design elements and the original, and proposes targeted optimisation strategies for urban park soundscape.