Music education psychology is the research on the changes in psychological activities in music teaching. It is the product of the combination of psychology and education. Besides, psychology can be adopted to study the changes in people’s psychological law in teaching. In cognitive psychology, the occurrence and defense of feelings such as feeling, attention, consciousness, knowledge, and gene can be systematically explained to provide a reference for people’s research on cognitive activities such as imagination, meaning, and thinking. The psychological activities related to cognitive psychology and sound provide a basis for studying new three-dimensional characters in national music teaching.

Music content and emotional expression are crucial contents running through the three links of creation, performance, and listening. Each link gives music special significance and vitality, and involves cognitive activities in people’s feelings, perceptions, and consciousness. Wu et al. (2018) studied the application of information and communication technology to traditional teaching to improve the effectiveness of teaching and training (Wu et al., 2018). Hence, the effect of introducing information technology and DL into the national music teaching classroom is analyzed by combining educational psychology to effectively make an accurate judgment on the application of intelligent music signal identification and generation technology in national music teaching (Wu and Song, 2019). Hence, psychology can be employed to analyze the psychological cognitive process of applying DL technology to national music teaching.

The most basic part of identifying music signals is the classification of music genres. Music is divided into multiple genres according to its characteristics, and the characteristics of different genres are also quite different (Jiang et al., 2020). However, music genres have similarities and differences (Kim and Oh, 2021). At present, the most widely recognized classification structure in the world mainly includes the GTZAN Genre and ISMIR2004 Genre. The GTZAN Genre mainly divides music genres into 10 categories: blues, country, hip-hop, jazz, pop, disco, classical, rock, reggae, and metal. The ISMIR2004 Genre mainly divides music into six genres: classical, electronic, jazz/blues, metal/punk, and rock/pop (Ng et al., 2020).

Music characteristic is the embodiment of the essential attribute of music. It is crucial to extract the characteristics of music to distinguish different music styles and genres (Caparrini et al., 2020). At present, there are two main types of music features: physical features and time-domain features. Short-term features are divided based on human sensory characteristics, mainly including tone, timbre, and loudness, which can be expressed by specific numerical features (Wick and Puppe, 2021). Some time-domain features cannot be expressed by specific numbers. The details are as follows:

(1) The short-time energy represents the amplitude of the music signal at a certain time. The calculation reads:

n represents the n-th sampling point, θ(m) is the signal value of the sampling point, ϕ(n−m) represents the window function, and A donates the window length (Wu and Wu, 2017).

(2) A crucial index to measure the high-frequency component of a signal is the short-time average cross zero ratios. In the waveform analysis, the more the high-frequency components are, the more the zero-crossing times are. Equation (2) donates this feature:

λ_n represents the short-time zero-crossing rate. θ(m) donates the signal value of the sampling point. ϕ(n−m) is the window function, and A is the window length; sgn represents the symbol function, that is, the value is 1 when x(m)≥0. Otherwise, the value is 0.

Audio storage formats mainly include MP3, Windows Media Audio (WMA), MIDI, and WaveForm (WAV). There are usually three types of events in music processing, MIDI events, system-specific events and meta events (Wu et al., 2019). The storage of audio is of great significance for music discrimination and generation (Siphocly et al., 2021). MIDI music is adopted here.

The most crucial thing in music signal identification is the identification of music style (Ramírez and Flores, 2020). The function of the model is to extract music features from different styles and genres of music in the music library through track separation technology, perform vectorization processing of the extracted music feature data, and then train the model. The LSTM network in DL is mainly adopted to enable computers to produce different styles of music (Sun et al., 2020). Figure 1 displays a flow chart.

Music data can be trained after processing (Hughes et al., 2018). In this model, track separation, music feature extraction, and data vectorization are mainly performed on the data (Gunawan et al., 2020). Figure 2 presents the architectural design of the model. MIDI music contains three data types, represented by 0, 1, and 2; 0 means that there is only one track; 1 means that multiple tracks will start playing in the same time series and at the same beat; 2 means that multiple tracks can be selected freely without starting simultaneously. Singletrack and independent multi-track MIDI files are relatively simple to extract. Since the tracks of independent multi-tracks are independent, the header files of the tracks need to be traversed in turn (Chen, 2019). All tracks of the music in format 1 are synchronized, so it is troublesome to separate, and the tracks need to be spliced. Music features are usually divided into physical features and time-domain features. Time-domain features can be only displayed by specific instruments (Rogoza et al., 2018). To better obtain the effective loudness, MIDI music extraction data containing at least 45 different loudness (the number of each loudness exceeds 20) are selected to become effective music data (Wu Y. J. et al., 2020). The effective music data are converted into a music score matrix as the input vector of the network. Meanwhile, the music score matrix needs to reflect the structure of MIDI and the characteristics of music in the form of vectors.

The music library contains music of various genres, and each genre has the same style. The system mainly needs to acquire the music characteristics of MIDI music in different genres, train the network, and then generate music of different styles (Zheng et al., 2018). The algorithm includes a music genre analysis model and a music style generation model. Figure 3 presents the model architecture. The music genre analysis model divides the learning problem into two parts. The first part is employed to learn the music features in the score and convert them into feature vectors, and the second part is to obtain the range of music intensity (Ahn et al., 2020).

Different genres of music have different musical characteristics. In this system, MIDI music of piano is mainly used. The main function of the music genre analysis model is to learn the style of music by learning a specific music genre, such as jazz, pop, and classical. (Yang and Lerch, 2020). Finally, different music genres will be distinguished according to the different music intensities. The music intensity matrix can be generated in the prediction part. The model mainly includes a bidirectional LSTM network layer and a linear layer (Hawley et al., 2020). The input of the model is a specific genre of music. Feature learning is carried out through a bidirectional LSTM network. The matrix containing sound intensity is generated through the linear layer. Finally, it is converted into music with the musical style that can be played (Chang et al., 2021).

The music sequences in the music library are different. Some music sequences are very long, while some are very short. At present, a bidirectional LSTM network has a good effect in dealing with long-time series problems (Briot, 2021). Bidirectional LSTM is more complex than unidirectional LSTM, which is mainly reflected in the value propagation process. Meanwhile, it needs more training times to optimize the parameters, while the unidirectional LSTM does not need multiple training times. The accuracy of training results of a bidirectional LSTM network is much higher than that of unidirectional LSTM (Jin et al., 2020) after many times of training. Figure 4 displays the LSTM network structure.

The main purpose of the model design is to generate music with a music style. As mentioned earlier, the final music genre style can be distinguished by the strength of music performance. Then, the music of different genres will have different strengths. The range of performance intensity is continuous and large, so it is essential to convert the output value into music strength value and change the output value range through the linear layer.

The music style analysis model is mainly adopted to learn more complex style information that cannot be trained in a music genre analysis network. The LSTM neural network is mainly used in this model. Figure 5 is the structural design of the music style analysis model.

The model is to study whether computers can learn and generate different music like people. The model mainly includes the interpretation layer and the subnet of the music genre analysis network. At present, the network is a multi-task learning model. MIDI music can be regarded as a music score. First, it passes through the interpretation layer, and then the output of the interpretation layer is adopted as the input of the music genre analysis network. After it passes through the music genres network, a matrix containing music characteristics will be output. Music analysis networks of different genres will generate matrixes of different genres. The matrix is converted into music that can be played (Shen et al., 2020).

The samples of music data to be analyzed are much fewer than other data, such as user information. The number of categories is relatively large, and there is no obvious distinction rule for each category. Hence, some scholars put forward the Siamese network, which is a similarity measurement method. It maps the input to the target space through a function and compares the similarity in the target space using Euclidean distance (Castillo and Flores, 2021).

In music style analysis, multiple units of the music genre analysis subnet need to be designed in the music style analysis model to better learn different music styles, because music has different genres. Each subnet is connected with the interpretation layer. The output of the interpretation layer is used as the input of the subnet unit. The music style analysis model includes multiple music genre analysis grids, and the biggest difference is an additional interpretation layer. The time of model training is reduced through the interpretation layer. The use of a multitasking mechanism can improve the efficiency of training and deal with the analysis of several kinds of music simultaneously.

A deep belief network combined with the Softmax algorithm is designed to classify the generated music genres. Equation (3) is the standard adopted to evaluate the accuracy of genre classification:

N represents the number of correctly identified music genres, M is the total number of music samples tested, and Q represents the accuracy.

The algorithm with the best classification effect among the music genre classification algorithms is selected in the experiment. The algorithm is based on the deep belief network in DL. After the network improvement, Softmax is adopted to predict the genre of music. RBM network belongs to a random network, and its particularity is mainly reflected in two aspects. The first is the probability distribution function. The node state of the network is random. The other is the energy function, which represents the stability of the network state. The greater the energy value is, the stabler the network state is. Equation (4) displays the definition equation of energy in the network.

ψ_i represents the value of neurons and w_ij donates the connection weight between neurons.

Iterations 1,000, 2,000, 3,000, 4,000, 5,000, and 6,000 are conducted in the experiment to find the best-generated music sequence. The error value of the experiment is increasingly smaller with the increasing iteration times of the LSTM network, which shows that the actual output value is increasingly closer to the target value, and the training results are increasingly accurate. Figure 6 displays the influence of iteration times on the training effect.

In the model design, the influence of the number of neurons in the hidden layer on the experiment is analyzed to optimize the model’s parameters. During the experiment, variables are set for the number of hidden layers and the corresponding number of neurons. The number of hidden layers is 1, 2, 3, and 4, and the number of neurons is 1,024, 512, 256, and 128, respectively. Figure 7 displays the effect of hidden layer neurons on experimental error. The number of hidden layers of the network is set to 4 layers, and the number of neurons in each layer is 1,024, 512, 256, and 128, which can minimize the optimal difference of the network training results.

In the experiment, different genres of music are processed by a fast Fourier transform. Then, the obtained music is analyzed by spectrum analysis and sound spectrum analysis to ensure the accuracy of the experimental data. The music sequence spectrum is analyzed during the experiment for the music sequence generated under different hidden layers. Figure 8 displays the generated music spectrum and sample spectrum. It shows that the effect of LSTM on music analysis is still obvious. The music spectrum at the training place is increasingly closer to the original spectrum with the increase of the number of the neural network layer, indicating that the accuracy is increasingly higher. Figure 8A displays the original music. Figure 8B shows that the learned music contains multiple unknown frequencies when there is only one hidden layer. Figure 8C displays that some frequencies do not appear when there are two layers. Figure 8D shows that the generated music sequence file is very similar to the original music sequence file with three layers, but there are still some differences. Figure 8E reveals that the difference between the generated music sequence and the original music sequence is very small when there are four layers, suggesting that the generated music is the most accurate with four hidden layers.

Overall, 5 genres of music are selected, and each music is in MIDI format, with 40 pieces of music of each type.

The classification accuracy of using a deep belief network combined with a Softmax algorithm is as follows (Table 1). The classification accuracy of jazz, classical, rock, country, and disco genres reaches 77.5, 65, 60, 67.5, and 70%, respectively. The analysis of experimental data reveals that the designed algorithm has better classification accuracy than the traditional algorithm. It shows that the music style and genre recognition and generation network in this experiment can generate music of different genres, and the accuracy rate is 60% or more.

The classification results show that the music style recognition and generation algorithm has a good performance in generating different genres of music. The spectrum of music generated by two different generation models: LSTM and RBM, are analyzed in this experiment.

The experimental results of different algorithms are compared in Figure 9. It reveals that the music generated by traditional RBM is the same as the original music, but the accuracy is not as high as that of the method proposed. Therefore, the designed algorithm can better generate different genres of music.