A two-way neural network music separation method for music intelligent classroom

Abstract

With the promotion of technology for educational reform and innovation, how to broaden the teaching space through technology and create a good classroom atmosphere in the music-smart classroom has become a hot topic for educators to explore. The study discusses music separation techniques based on those commonly used in the intelligent classroom. To address the problem of using the sample timing information in the training process, the study uses LSTM networks instead of traditional recurrent neural networks. It constructs a DS_BRNN algorithm for the separation of accompaniment and song of mixed music. A discriminative training objective function is introduced to train the real part separately from the imaginary part, aiming to extend the separation target from the real domain amplitude spectrum to the complex domain amplitude spectrum. The innovation of this research lies in using the single-channel music separation method to improve the teaching effect of music intelligent classrooms. The results on accompaniment separation performance showed that the DS-BRNN algorithm was 0.161 dB lower than the DNN music separation model in GSAR values but improved by about 2.5–4.3 dB in GSIR and GSDR values. Moreover, it also had a similar performance in separating human voices, while the GSIR value of HPSS was only about 3 dB higher than that of DS-BRNN. The proposed improved algorithm has better comprehensive performance than other traditional separation models in music separation. The primary contribution is to provide technical support for the intelligentization of music classrooms and to establish a theoretical basis and potential applications for the creation of teaching situations that utilize music separation in intelligent music classrooms.