Polyphonic choral singing can not only cultivate musical imagination and improve musical literacy but also allow singers to feel the harmonious and beautiful musical rhythm during polyphonic choral singing. To improve the accuracy of polyphonic singing, the study designed a music source separation structure based on recurrent neural networks using deep learning technology. And, combined with ResNet and CBAM, a joint neural network based on Res-CBAM was designed for optimization. After that, the main melody of the human voice was extracted using the polyphonic music melody extraction algorithm that was created in this paper. The listening training was then done in three areas: pitch, rhythmic rhythm, and vocal balance. The trained community choir members showed significant improvements in singing ability, breath control, pitch, rhythm, polyphonic choral ability, and expressiveness (p<0.05). It indicates that the auditory discrimination training based on the polyphonic music melody extraction algorithm has a facilitative effect on the accuracy of polyphonic singing in the choir.