<p>Due to the extremely shortage of labeled data in low-resource scenario, the deep learning models are easily overfitted and extracted features failed in English dialect recognition tasks. In this article, curvature-aware gradient projection prototype meta-learning framework is proposed. This framework constructs a Riemannian metric space with second order curvature information and utilizes geodesic distance to improve the geometric representation of subtle acoustic variation of dialects. In backpropagation, eigenvalues of local Hessian matrix of loss function are dynamically calculated, and gradient update direction is orthogonally projected into flat subspace of loss curvature to suppress noise interference, thereby realizing rapid adaptive transfer of general parameters to new dialect categories. Experiment results show that the method achieves recognition accuracy of 78.4% under highly challenging 5-way 1-shot sample setting, and the average accuracy of 87.6% under circumstance of facing unseen English dialect variants. The method has confirmed that the meta-learning method containing curvature-aware mechanism can solve the problem of data shortage in low-resource dialect recognition, and provides a new technical route for fine-grained speech classification.</p>