Singing in online singing platforms has become popular nowadays. An accurate singing-song recommendation system is essential for such platforms to help users ?nd desirable and suitable songs to sing and thus enhances their experience and stickiness. One common idea behind most existing singing-song recommendation methods is to recommend songs that match a user’s vocal competence, i.e., the user’s capability of singing. However, e?ective and e?cient way to estimate user’s vocal competence is still needed. Meanwhile, users in online singing platforms can easily get access to the published and reposted recordings of their online friends, which may in?uence users’ singing behaviors. How to exploit the information related to publishing, reposting and social interaction behaviors of users’ online friends for recommendation is essential but not well studied in literature. In this paper, we study how to solve these issues. We propose an automatic method to measure a user’s vocal competence on a song. Meanwhile, we propose a graph convolutional neural network model that leverages the behavior information of a user’s online friends to model social in?uence on the user’s singing behavior. Finally, we develop an integrated model for singing-song recommendation, taking both vocal competence and social in?uence into consideration and train the model by an end-to-end way. Experiments conducted on a real-world dataset demonstrate that our proposed model can improve recommendation performance substantially and outperforms other baseline methods.