Steady-state visual evoked potentials (SSVEP)-based brain–computer interfaces (BCIs) leverage machine learning methods to enhance performance. However, these methods require a sufficiently long time window to achieve high accuracy and information transfer rate (ITR), which restricts their applications in real-world scenarios, particularly for user-specific decoding. To address this issue, we propose a parameter-based transfer learning CNN (PTL-CNN) approach for the SSVEP-BCI system, which can automatically fuse and extract both inter- and intra-subject features in EEG signals. Specifically, we first introduce a shallow CNN architecture and adopt a short time-window to train a pretrained model on a dataset comprising numerous subjects, aiming to explore the universal features across subjects. Subsequently, a new user is utilized to fine-tune the model, calibrating it to this specific user. Experimental results demonstrate that PTL-CNN achieves remarkable performance and significantly outperforms the compared algorithms under short time windows. For instance, in a time window of 0.4 s, PTL-CNN achieves an average accuracy of 80.60% with an average ITR of 247.77 bits/min on the Benchmark dataset, and an average accuracy of 66.91% with an average ITR of 185.90 bits/min on the Beta dataset. This performance is significantly better than that of Ensemble-TRCA (Benchmark: 71.21%, 209.12 bits/min; Beta: 53.04%, 135.53 bits/min). In summary, our proposed PTL-CNN achieves the highest average accuracy with the fastest average ITR and is of implications for the development of real-time BCI applications, as well as inspiration for other application paradigms.