The potential of motor imagery-based brain-computer interfaces (BCIs) is hindered by long calibration times. Therefore, this study investigates a classification model that minimises BCI calibration time while maximising its accuracy by exploiting transfer learning. To this end, a modified version of the Sinc-EEGNet architecture is proposed. Analyses were carried out with data from multiple subjects. Notably, when the model was trained with data from subjects other than the test subject, Sine-EEGNet-32 achieved a mean classification accuracy of 78 $\pm$ 10 %. This outperformed the reference EEGNet-4 architecture by 10 %. Instead, when considering also data from the test subject for a fine tuning, Sinc-EEGNet-32 achieved a mean accuracy of 80 $\pm10\ \%$ by exploiting only 10 % of test subject's data and 83 $+10\ \%$ by exploiting 40 % of test subject's data. These correspond to a system calibration of less than 2.0 min and of approximately 8.0 min, respectively. Overall, there was an increasing trend in performance for Sinc-EEGNet-32 as higher percentages of data were exploited for fine-tuning. In contrast, EEGNet-4 only achieved an accuracy of 72 $\pm$ 13 % even with fine tuning.