Impact-echo is one extensively applied non-destructive technique for flaw detection in concrete structures. In impact-echo test, surface motion generated by short-duration mechanical impact is investigated for structural condition assessment. This paper endeavours to formulate impact echo analysis by using novel statistical techniques, i.e. Grassmann manifold learning. Comparing to conventional impact-echo test, the proposed method presents several favourable properties: 1. Conventional impact-echo method mostly relies on frequency peak in echo spectrum, the proposed method characterizes rich temporal-spectral patterns in addition to the spectral peak. 2. Proposed method is performed over local area on concrete surface with integration of several consecutive echo responses, and thus produces more stable condition evaluation result comparing to point-wise impact-echo approach. 3. To cope with extracted echo feature, effective similarity metric on Grassmann manifold is employed, which favourably facilities condition-based assessment. To demonstrate the proposed method, we prepared concrete specimen with 2cm, 4cm and 6cm depth void inside and echo signal is captured through air-coupled sensor. Experimental result demonstrates the effectiveness of the proposed method, including accurate condition-based classification performance and high processing efficiency