Pointing at the questions of subjective estimation, fragile handing-down and unbalanced teaching in traditional local opera passing-on, this thesis puts forward an intelligent identifying and teaching estimating system which is based on CNN-LSTM-Attention. This system carries out the integration of multi-layer acoustic feature extraction and time sequence modeling, for the identification of opera singing styles and the evaluation of technical, artistic and cultural authenticity dimensions. The experiments which we carry out on the self-constructed LOFRS data set (5,240 sample pieces, 131.1 hours) indicate that the recognition accurate rate achieves 94.2%, with 18ms inference delay time and 14.7MB model dimension. A 16-week teaching experiment that includes 100 students has proven that this system can significantly promote learning results (p<0.001, Cohen’s d=1.25–1.78). This research gives an effective and uniform technical plan for the intelligent passing-down and individual guiding teaching of opera art.
