There is an abundance of materials for use in professional music courses, but it can be difficult for consumers to quickly and efficiently obtain the specific knowledge they require. Additionally, there is a general lack of data collected on online learning, which renders the recommendation effect of music course resources insufficient. In this study, we use technologies connected to the knowledge graph to the field of online education in order to create a system capable of recommending acceptable educational materials for use in professional music classes. In order to construct a recommendation model using multi-task feature learning, knowledge graphs are embedded within tasks, and high-order connections between latent features and entities are constructed across tasks using cross-compression units. It is possible to achieve success by recommending relevant course materials for individual students based on their requirements, interests, and present skill levels. In terms of its ability to generate suggestions, the proposed knowledge spectrogram-based teaching resource recommendation system for professional music courses outperforms four baseline models on a number of publicly available datasets. This method has some practical utility in the domain of course resource suggestion, and its training time is less than that of the comparison model.