Deep Learning-Based DOA estimation commonly suffers from grid quantization errors in classification methods or fixed-output constraints in regression methods. This letter proposes Spectral Implicit Neural Representations, termed SPINRs, an arbitrary-angle querying framework that decouples the estimator into a backbone-agnostic Sample Covariance Matrix encoder and an angle-conditioned scalar spectrum decoder. The encoder extracts global spatial context, while the decoder learns a smooth mapping from angular coordinates and observation features to spatial-spectrum values. Because the angle is supplied as an input rather than represented by a fixed output neuron, the queried coordinate set can be adjusted without modifying the output layer or retraining the network, and the number of trainable parameters is independent of query density. The same decoder interface can also be attached to different feature backbones. Experiments with CNN, ResNet, and UNet encoders demonstrate improved off-grid localization relative to their fixed-grid counterparts under practical finite-SNR conditions, especially in low-SNR regimes.
