Deep Learning Algorithm Based Pattern Recognition of Musical Melodies and Its Effectiveness Optimization in Education

In order to solve the shortcomings of the sound source separation method, this paper proposes a melody extraction method based on saliency and improved joint neural network, constructs the pitch saliency feature function according to the idea of harmonic energy superposition, pre-processes the audio, and then builds the joint neural network based on Res-CBAM according to the idea of joint neural network of music detection and pitch estimation classification to realize the melody pitch contour tracking. In addition, the calculation of the significance function is introduced to highlight the pitch significance features, so that the graphs input to the neural network have clearer melodic features. The results show that before and after the suppression of the accompaniment, the difference in the time-domain waveforms is not significant in the treble range, but there is a significant difference in the low-frequency range. In addition, the OA accuracy of the Res-CBAM algorithm proposed in this paper is up to 41.14% higher than other algorithms (P < 0.05), and the accuracy of the model is good. Applying this recognition model to teaching found that teaching with this model can significantly improve the subjects' perception of music (t=.197, p=0.002<0.05). It can be seen that the application of the Res-CBAM algorithm to actual music teaching is of great practical importance.