The modelling of coastal Directional Wave Spectra (DWSs) often requires downscaling techniques integrating DWSs from open ocean boundaries. Dynamic downscaling methods reliant on numerical wave models are often computationally expensive. In coastal areas, wave dynamics are strongly influenced by the topography, implying that once the DWSs at the open ocean boundary are known, the DWSs at various locations along the coast are almost determined. This property can be utilized for statistical downscaling of coastal DWSs. This study presents a deep learning approach that can compute coastal DWSs from open ocean DWSs. The performance of the proposed downscaling model was evaluated using both numerical wave model data and buoy data in the Southern California Bight. The results show that the deep learning approach can effectively and efficiently downscale coastal DWSs without relying on any predefined spectral shapes, thereby holding promise for coastal wave climate studies.