In this paper, we explore a new method for skeleton-based human action recognition. First, the normalized angles of local joints are extracted and Spherical Harmonics Transform (SHT) can then be used to explicitly model the angular skeleton by projecting the spherical angles onto unit sphere basis. This enables that the skeleton representation can be decomposed into a basis functions. We adopt the spatiotemporal covariance matrix of the spherical harmonic to capture joints orientations over the human action sequence. Thus, the co-variance coefficients of joints are used as a discriminative de-scriptor for the sequence. We validate the proposed method using Extreme Learning Machine (ELM) classifier and recent published 3D action datasets. Experimental results show that our method performs better than many classical methods.