Surface roughness is an essential attribute of surface quality in the machining process to ensure the mechanical functionality of machined parts. Effective in-process monitoring of surface roughness is indispensable, which can be achieved through intelligent monitoring and prediction systems, utilizing data-driven models and dynamic process signatures. Notably, the performance of these systems, particularly in terms of prediction accuracy and robustness, is hindered by the limited application of data-driven models, especially deep learning models, which are intensive data-demanding. The inaccessibility of ample surface roughness data and corresponding process signatures leveraged for model building poses a significant challenge. In response to this challenge, as an emerging generative modelling approach, a classifier guidance diffusion model renowned for process stability and image generation with high fidelity and sufficient diversity is introduced to surmount data scarcity. The pre- trained diffusion is applied to synthesize audible sound signals collected from a machining process with the guidance of gradients from a classifier. After that, ground truth signal data mixed with synthesized ones in different proportions are leveraged to train and test the pre-trained model – VGG16, for surface roughness prediction in a transfer learning process. The results demonstrate that the proposed model can effectively generate synthesized 2D Mel-spectrogram images of audible sound signals with superior quality, further boosting the prediction performance of VGG16 in terms of its prediction accuracy and generalization under an enriched dataset.