The grating lobe appears when the microphone array element spacing is larger than the half-wavelength of the acoustic signal for the DOA estimation of the wideband sound source. Although the utilization of multi-frequency bins data can suppress the grating lobes to some degree, the current methods perform unsatisfactorily and are computationally inefficient. To address these issues, an improved method based on the sparse Bayesian learning is proposed for wideband DOA estimation. This method introduces the hyperprior to the multi-frequency sparse Bayesian estimation model, and then takes advantage of the fact that the source signal has the same sparsity in multi-frequency bins and combines the expectation maximization algorithm to derive the iterative form of each hyperparameter in the multi-frequency sparse Bayesian model. In addition, an off-grid model for the wideband sound source is incorporated into the proposed framework to better fit the practical scenarios. To evaluate the performance of the algorithm, simulations and field experiments are implemented. Results show that the proposed method can better exploit the multi-frequency characteristics of the wideband sound source to reduce the interference of the grating lobes, while having higher estimation accuracy and faster estimation speed compared with the multi-frequency compressive sensing method with l 1 minimization and the multi-frequency sparse Bayesian learning method. In the practical tests, the improved method shows better grating lobe suppression ability than other advanced methods, and its estimation error can reach 0. 09°. Compared with MF-SBL, the number of iterative convergence steps required is reduced by about 50% .