Intelligent Identification and Algorithmic Optimization of Chinese Traditional Music Elements in Dance Performance in the Internet Era

There are many forms of music recognition, but because they are based on the subject of music, they are bound to follow the internal principles of music and are built on the basis of the form of music, i.e., the acoustic structure of music. The form of music consists of the basic elements: pitch, duration, intensity and quality. These elements are combined to produce the means of organization of music: rhythm, melody, and harmony. And through this organization, we can further explore the structure, style, and emotion of music. The three levels of music: the basic elements, the organizing means, and the compositional structure present a progressive organizational relationship between point-line-plane and plane.

In music theory, there is a category of music that considers that the organization of the elements within the acoustic structure of music is closely related to the proportion and arrangement of numbers. “Form is the organization or arrangement of the various elements that make up a thing”, and the form of music consists of three aspects: the basic elements, the means of organization, and the medium between the basic elements and the means of organization - the law of beauty of form. The basic elements include tempo, intensity, intervals and rhythm. Organizational means include harmony, melody, tune, modulation, etc. And the form is a synthesis of musical forms, the way of combining elements such as melody, rhythm, beat, harmony, polyphony, modulation, tonality, timbre, tempo, intensity, strength, and the structural layout and development program. The form of music has relative independence and a positive dynamic role, and it is continuous, irreversible, independent, and relative. The form of music is the method and means of expression for the musical language, with its own specific organization and structure.

The types of music will be classified into six categories according to the way of recognition: melody recognition, rhythm recognition, harmony recognition, song structure recognition, music style recognition, and music emotion recognition. Since the classification is made according to the way of application of machine recognition, different forms of machine language will be used as reference materials according to the conversion, and the classification means are as follows: in the two categories of melodic recognition and rhythmic recognition, since both of them are the organization of musical elements, they will be reclassified according to the combination of the elements, e.g., the melody is the pitch and the rhythm is the pitch and the rhythm is the tone and the tone strength.

Artificial intelligence has been committed to making computers have “human ears”, so that machines can simulate human feelings, which makes it necessary to convert the music language into a language that can be recognized by computers, and from the point of view of the different means of the elements of music in this paper, the way of transformation and the form of the transformation are different for different levels.

In terms of music style detection only the overall characteristics of the music can be judged. In terms of music style creation, it is possible to compose according to the characteristics of a certain type of music.

The traditional national tuning recognition method is shown in Figure 1, where the national tuning is based on pentatonic tuning, and the five tones with the highest frequency of use are used as the basic level of the piece, and less than five tones are all basic levels. After setting the confidence function to extract the tone with the largest index as the main tone, the lowest pitch tone among the three tones constituting two consecutive major diatonic relationships is selected as the palace tone.

Figure 1.

Traditional national mode recognition method

The current techniques for analyzing musical structures can be divided into two parts: “states” and “sequences”. “States” are sets of near-times that contain the same sound information, and the audio signal is viewed as a series of states at different scales. A series of states of different scales, corresponding to the different structural types of a song, closely related to the structure of the overture, the main song, the chorus, will be calculated through the pitch, spectral and fragment characteristics, but only limited to music with high repetition. “Sequences”, i.e., collections of consecutive times that are similar to collections of consecutive times, are closely related to chords and melodies. It is generally based on the separation of spectral shapes, the cutting of overtones and the division of rhythmic pitches. Both are mainly applied to popular music. Recognition of music emotion is mainly based on the characterization of music using emotion modeling.

The basic elements of music and their inter-compositional features such as pitch variation, rhythmic speed, etc. They are matched with manually labeled music emotion types and trained to recognize the emotion.

Corresponding to the manually labeled types of musical emotions, training the machine to finally accomplish the goal of recognizing the emotions.

Because the differences of traditional music genres are mainly reflected in the differences of melody and rhythm, the features that characterize these musical properties should be long-lasting. Therefore, in this paper, the short-time features are grouped into long-time features, i.e., the 480-dimensional MPC parameters are used as the input vectors of the network. Input to the deep confidence network. The original features are inputted from the input layer and reach the output layer through multiple hidden layers. Each hidden layer decomposes and reconstructs the feature vector of the previous layer. In turn, the original features are transformed into new forms of feature representations, and the newly obtained abstract features are used to identify the category of each test music sample through the output layer (Sofimax layer, i.e., classifier).

Each layer of the DBN contains multiple neural units, and the state of each neural unit node is jointly determined by the states of the other nodes and their connection weights to this node multiplied and combined with an activation function. Pre-emphasis is generally realized by a first-order digital filter before feature parameter extraction, and the transfer function of this filter is expressed as: (1)H(z)=1−az−1$$H(z) = 1 - az^{-1}$$

Where, a is the pre-emphasis factor, which is generally taken as a decimal close to 1. Let the sampling value of the genre music signal at moment n be. C toward, then the result after the pre-emphasis processing is: (2)y(n)=x(n)−ax(n−1)$$y(n) = x(n) - ax(n-1)$$

The theoretical formula for the number of strands of a music signal fragment is: (3)N=[N1−N0N2−N0]$$N = [ \frac{N_{1-}RV N_0}{N_{2-}N_0}]$$

In general, the inputs to the neurons are 0-1 binary, but since the MPC features of the experimental input network in this paper are real, I set the first RBM as a Gauss-Bernoulli constrained Boltzmann machine structure with activation probabilities of the nodes in the hidden and visible layers, respectively: (4)p(hj=1|v)=σ(at+∑i=1mviwij)$$p\left( {{h_j} = 1|v} \right) = \sigma \left( {{a_t} + \sum\limits_{i = 1}^m {{v_i}} {w_{ij}}} \right)$$ (5)p(vi=1|h)=N(ht+∑i=1nhiwij,1)$$p\left( {{v_i} = 1|h} \right) = N\left( {{h_t} + \sum\limits_{i = 1}^n {{h_i}} {w_{ij}},1} \right)$$

where σ(x)$$\sigma \left( x \right)$$ is the sigmoid function and N(x)$$N\left( x \right)$$ is the normal distribution function with mean 0 and variance 1.

So the MPC feature parameters of the music samples of each genre also need to be standardized before they are input into the DBN.

At the same time, in order to prevent the occurrence of large prediction errors due to the order of magnitude difference of the data in each dimension, the original feature parameters of all the manually extracted samples are standardized by dimension, so that the feature parameters in each dimension obey the normal distribution with mean 0 and variance 1, and the standardization formula is: (6)Standardizeddata=Rawdata−meanStandardDeviation$$S\tan dardized{\text{ }}data = \frac{{Raw{\text{ }}data - mean}}{{S\tan dard{\text{ }}Deviation}}$$

In addition, since the neuron inputs to the intermediate hidden layer are 0-1 binary, the remaining three RBMs are structured as Bernoulli-Bernoulli constrained Boltzmann machines.

The training process of DBN is generally as follows: first, randomly initialize the weights and biases of the network, then sequentially and unsupervised pre-train each RBM, and when all the RBMs have been pre-trained, then assign the weights and biases obtained from the training to the back-propagation network, and achieve the fine-tuning of weights and biases to the whole network by adjusting the error between the actual labels of the music samples and the predicted labels.

Subtracting the activation probability of the original visible layer v with the activation probability of the reconstructed visible layer v, the result obtained is regarded as the increment of the visible layer v corresponding to the bias b. Subtracting the activation probability of the obtained hidden layer h and the activation probability of the hidden layer h, the result obtained is regarded as the increment of the hidden layer h corresponding to the bias a. Subtracting with the obtained probability of forward propagation and the probability of back propagation, the obtained result is regarded as the increment of the weight matrix W. And during each iteration, the updating of weights and biases is performed at the same time, so it converges at the same time. Combined with the corresponding learning rate, the weights and biases are updated according to Eqs. (7) to (9): (7)Wt=Wt−1+ε(vp(h|v)−v′p(h′|v′))$${W_t} = {W_{t - 1}} + \varepsilon \left( {vp(h|v) - v\prime p\left( {h\prime |v\prime } \right)} \right)$$ (8)bt=bt−1+ε(p(v|h)−p(v′|h′))$${b_t} = {b_{t - 1}} + \varepsilon \left( {p(v|h) - p\left( {v\prime |h\prime } \right)} \right)$$ (9)at=at−1+ε(p(h|v)−p(h′|v′))$${a_t} = {a_{t - 1}} + \varepsilon \left( {p(h|v) - p\left( {h\prime |v\prime } \right)} \right)$$

When training a deep neural network, if the training samples are insufficient, the network tends to overfit, which leads to a poor generalization ability of the model. Since the training process of RBM, the constituent unit of DBN, itself has some antioverfitting ability, the DBN algorithm used in the previous subsection has good model results even without the additional antioverfitting method.

In this subsection, an improved DBN algorithm is proposed to improve the performance of music genre recognition and classification. This improved DBN algorithm refers to embedding a Dropout layer between the visible and hidden layers of each RBM in the DBN to further prevent overfitting, which effectively improves the generalization ability of the multilayer neural network, and introduces momentum during the update of the weights in each layer to prevent the network from falling into local minima.

In order to improve the professionalism of this study, and committed to obtaining a more objective and scientific analysis process, this study is aimed at 10 experts in traditional dance performance-related research to conduct expert interviews, especially for the dance performance in the creation of a few elements need to pay attention to ask questions to achieve an in-depth analysis of synergistic elements of the creation of the dance performance. Preliminary factor indicators were obtained, which were generally categorized into three major categories: music, movement, and function, as shown in Table 1.

We can see that after the second and third rounds of organizing the factors and expert selection of the factors, there are a total of 11 factors in the 3 major categories of factors. There are 10 experts’ choices for music rhythm, 7 experts’ choices for music melody and 7 experts’ choices for music emotion in the factors of music elements. There were 7 expert choices for action parts, 6 expert choices for action forms, 6 expert choices for action complexity, 5 expert choices for space utilization, and 4 expert choices for action styles in the factor of action elements. The factor of functional elements was chosen by 6 experts for fitness effect, 7 experts for aesthetic emotion and 4 experts for traditional clothing. From these, we selected the factors chosen by more than half of the experts were 6 factors of music rhythm, music melody, movement parts, movement form, fitness effect and aesthetic emotion for the next analysis.

In order to better analyze the weights of the six factors identified above, a questionnaire was distributed to score the weights of the six elemental factors above, so as to explore in depth the different weights of different groups of people’s views on the synergistic factors of dance performance. Ten dance performance-related experts, social professionals, and social ordinary participants, totaling 50 people, were chosen to be interviewed in this paper. The results of the survey are shown in Table 2, and the overall scores for the six factors were summarized by conducting a fourth round of weighted scoring. The total score for music rhythm was 219, for music melody was 216, for movement parts was 206, for movement form was 193, for fitness effect was 194, and for aesthetic emotion was 186. It can be seen that the first factor is the rhythm and melody of the music. The second factor is the movement parts and fitness effect, followed by aesthetic emotion and movement form.

The choreographic factors of traditional dance performances are shown in Table 3. In the Dai dance performance, “South of Colorful Clouds” adopts 3/4 beat in terms of rhythm, with a total of 2 segments and a duration of 2 minutes and 54 seconds. The Moon” adopts 3/4 beat in terms of rhythm, and also has 2 sections, with a duration of 3 minutes and 6 seconds. The music of the two dance performances is slightly different in terms of tempo. The music of South of Colorful Clouds has a more brisk tempo, while the tempo of The Moon is relatively calmer. The two songs have the same number of passages, but there is a slight difference in length, a difference of 12 seconds, the length of “The Moon” is more than the length of “South of Colorful Clouds”. It can be seen that the overall rhythm of the dance performance of “The Moon” is slower than that of “South of Colorful Clouds”.

The Tibetan Dance Performance “Ballad of Auspiciousness” is in 2/4 time, with 2 sections and a duration of 3 minutes and 16 seconds. The rhythm of “Heavenly Love” is in 3/4 time, also with 2 sections, and the duration is 3 minutes and 45 seconds. The music of the two sets of dance performances is the same in terms of rhythm and the number of paragraphs, but there is a difference of 29 seconds, nearly half a minute, in terms of duration, and the duration of “Heavenly Love” is longer. Therefore, it can be seen that the practicing duration of each paragraph of “Heavenly Love” will be longer in the two sets of Tibetan dance performances.

The rhythmic aspect of “Waiting for You to Come to the Grassland” in the Mongolian Dance Performance is in 4/4 time, with 3 sections, and the duration is 3 minutes and 42 seconds. The rhythm of “Searching for the Heart” is in 3/4 time, with 2 sections and a duration of 4 minutes and 16 seconds. The music of the two sets of dance performances is slightly different in terms of rhythm, the music of “Waiting for You to Come to the Grassland” has a more distinct and brisk rhythm, while the music of “Seeking of the Heart” has a more gentle rhythm. The two songs have the same number of passages, but the difference in duration is 34 seconds, which is a big difference. Therefore, it can be seen that the overall tempo of “Heart’s Search” is slower than that of “Waiting for You to Come to the Grassland” in both sets of Mongolian dance performances.

Uyghur Dance Performance “Lift Up Your Head” is in 3/4 time, with 3 sections and a duration of 2 minutes 32 seconds. Why Are the Flowers So Red” is in 2/4 time, with 3 sections, and the duration is 3 minutes 54 seconds. The two sets of Uyghur dance performances have the same rhythm in the music and slightly different paragraphs and durations. It can be seen that “Lift Up Your Cover” has one more paragraph than “Why Are the Flowers So Red”, but the durations are not as long as that of “Why Are the Flowers So Red”, and the difference between the two sets of Uyghur dance performances is 1:22, which is a big difference. It can be seen that in the two sets of Uyghur dance performances, “Lift Your Cover” has a shorter length of practice per section, and “Why Are Flowers So Red” has a longer length of practice per section.

In this section, the improved DBN algorithm is used to identify and classify the musical elements, tunes, emotions and themes in the dance performance, to express different regional characteristics and traditional cultures with different timbres, to accentuate the theme of the dance, and to resonate with the audience. Traditional music shows the artistic feeling of Chinese traditional music through the most traditional Chinese forms and singing methods and tunes, accompanied by different forms of opera singing, bringing different styles to the dance and giving the audience a unique flavor.