Automatic singing evaluation is highly desirable in industrial performance evaluation scenarios, including education and entertainment, given the high cost of human judges. Current singing evaluation systems assess a singer's performance by comparing reference vocals to the singing track. Unfortunately, reference vocals are not always available. Moreover, the similarity measure may not provide a reliable evaluation due to the inherently variable nature of singing performances. This paper proposes a tonality-based method for song-independent automatic singing evaluation. Experimental results show the proposed method outperforms other non-intrusive evaluation algorithms.