Abstract Online teaching is becoming an important means of education, but for the piano course, how to organically combine online and offline teaching has been the focus of educators. To improve this situation, this paper proposes a neural network model for piano performance evaluation as an auxiliary means of blended teaching. The model uses bidirectional LSTM training data, introduces an attention mechanism to optimize data recognition efficiency, and finally generates performance evaluation through a softmax classification algorithm. After implementing online-offline hybrid piano teaching in combination with this evaluation model, 66.72% of the students believed that the new teaching method could stimulate learning initiatives. The piano performance test scores showed that the post-test times for single-note playing, mixed intervals and chords, and two-handed ensemble playing in the experimental group were reduced by 38 seconds, 45 seconds, and 7 minutes, respectively, and were significantly higher than those in the control group. This indicates that the performance evaluation model has good performance and application effectiveness.