KR20170128073A - Music composition method based on deep reinforcement learning - Google Patents

Abstract

The present invention provides a method for composing a music using an artificial intelligence module considering all of a rhythm, a harmony, and a melody. The music composition method using deep reinforcement learning comprises the following steps of: a computer device inputting a seed melody into a melody neural network in which the melody is learned and extracting a constant melody line; the computer device inputting the seed melody into a rhythm neural network in which the rhythm is learned and extracting a constant rhythm; and the computer device considering the rhythm over time flow and adjusting the melody line.

Description

[0001] MUSIC COMPOSITION METHOD BASED ON DEEP REINFORCEMENT LEARNING [0002]

The techniques described below are directed to techniques for composing music using in-depth neural networks and reinforcement learning strategies.

Recently, there has been an attempt to automatically compose music using artificial intelligence. Google has introduced artificial intelligence to compose music through the Magenta project, while Sony has introduced an artificial intelligence called Deep Bach.

On the other hand, the interface for easy composing by non-experts is being studied. For example, when a user paints a bar-shaped space by providing one or more bar displays corresponding to the piano keys, a user interface is disclosed in which a key corresponding to the bar is reproduced for a time corresponding to the length of the colored box .

Korean Patent Publication No. 2012-0060085 (2012.06.11)

Google's magenta is based on listing chords progressively and listing bits based on that chord. A melody is a variation in which three or four of the four basic tones that make up each chord are picked up and simply stacked up from the bottom up to match the tune of the song. The result of magenta is lack of musical perfection due to simple melody repetition. For example, magenta melody is actually a half-note music with only harmony and accompaniment, since it is nothing more than a straight line of musical notes corresponding to the left hand accompaniment of the piano.

On the other hand, Sony's Deep Bach is characterized by studying Bach's music in a deep-running manner and composing a choral song with a distinctive atmosphere of Bach. Therefore, Deepbach can only produce songs with a limited atmosphere due to the limitation of training data. Deep Bach is only able to fulfill the musical elements composed of Mars, accompaniment, rhythm and melody, since it provides the sound of a four-member ensemble at every moment.

The technique described below is intended to provide a method of composing music using an artificial intelligence model considering both rhythm, harmony, and melody. Further, the technology described below is intended to provide an interface through which a user, not an expert, can easily compose music.

A method of music composition using deep reinforcement learning comprises the steps of: inputting a seed melody to a melody neural network in which a melody is learned, and extracting a certain melody line; and the computer device inputs rhythm of a rhythm into a rhythm- And adjusting the melody line in consideration of the rhythm according to the flow of time.

Using the technique described below, a user can easily compose beautiful music with harmonious rhythm, harmony and melody.

1 is an example of a process of composing music using a modularized neural network.
2 shows an example of a neural network model for music composition.
3 shows an example of the structure of a neural network module.
4 is another example showing a neural network model for music composition.

The following description is intended to illustrate and describe specific embodiments in the drawings, since various changes may be made and the embodiments may have various embodiments. However, it should be understood that the following description does not limit the specific embodiments, but includes all changes, equivalents, and alternatives falling within the spirit and scope of the following description.

The terms first, second, A, B, etc., may be used to describe various components, but the components are not limited by the terms, but may be used to distinguish one component from another . For example, without departing from the scope of the following description, the first component may be referred to as a second component, and similarly, the second component may also be referred to as a first component. And / or < / RTI > includes any combination of a plurality of related listed items or any of a plurality of related listed items.

As used herein, the singular " include "should be understood to include a plurality of representations unless the context clearly dictates otherwise, and the terms" comprises & , Parts or combinations thereof, and does not preclude the presence or addition of one or more other features, integers, steps, components, components, or combinations thereof.

Before describing the drawings in detail, it is to be clarified that the division of constituent parts in this specification is merely a division by main functions of each constituent part. That is, two or more constituent parts to be described below may be combined into one constituent part, or one constituent part may be divided into two or more functions according to functions that are more subdivided. In addition, each of the constituent units described below may additionally perform some or all of the functions of other constituent units in addition to the main functions of the constituent units themselves, and that some of the main functions, And may be carried out in a dedicated manner.

Also, in performing a method or an operation method, each of the processes constituting the method may take place differently from the stated order unless clearly specified in the context. That is, each process may occur in the same order as described, may be performed substantially concurrently, or may be performed in the opposite order.

The techniques described below compose music using a plurality of artificial intelligence models that sequentially arrange musical elements based on reinforcement learning. In the following description, it is assumed that a computer device composes music using an artificial intelligence model. A computer device refers to a smart device, a PC, a server on a network, and the like.

1 is an example of a process of composing music using a modularized neural network. The user composes music using the computer device 100. [ The user may enter the seed element into the computer device 100. [ The seed element is input to the composition artificial intelligence model and used to generate certain music. The seed element corresponds to start information for composition, and may include information such as a certain rhythm, a melody, and a chord.

Further, the user can input certain information for creating a seed element in the computer device 100, such as a smart device, as shown in FIG. For example, a user may input a line, which is a graphical element, through a touch panel (or other interface device). As shown in Fig. 1, the user can input a line having a certain change. The types of graphic objects input by the user may vary. The user can also input lines, points, curves, specific shapes, and the like. Furthermore, the user can directly input voice or music to the computer device. The computer device can extract the seed element based on the information input by the user using a separate artificial intelligence model.

2 shows an example of a neural network model for music composition. 2 is an example of an artificial intelligence model for generating music. The computer device can compose music using a plurality of neural networks as shown in FIG. For example, a neural network model outputting a code, a neural network model outputting a melody, and a neural network model outputting a rhythm can be used. That is, the computer device can compose music using a plurality of modularized neural networks.

The code neural network 30 is learned using sample codes in advance. The code neural network 30 receives the seed code and outputs a predetermined code. The seed code is one of the seed elements described above.

The melody neural network 50 is learned using a sample melody in advance. The melody neural network 50 receives the seed melody and outputs a certain melody. Seed melody is one of the seed elements mentioned above.

The rhythmic neural network 80 is learned using a sample rhythm in advance. The rhythm neural network 80 receives the rhythm of the seed and outputs a predetermined rhythm. Seed rhythm is one of the seed elements described above.

The post-processing neural network 90 receives the code, the melody, and the rhythm, and outputs the final music. The post-processing neural network 90 can modify the melody in consideration of the code progress. For example, post-processing neural network 90 may remove certain notes in the melody line or adjust the melody line's tone to match the composition with the chord. The post-processing neural network 90 may adjust the position or length of the melody line to match the rhythm.

The code neural network 30, the melody neural network 50, the rhythmic neural network 80, and the post-processing neural network 90 use the reinforcement learning method. The code neural network 30, the melody neural network 50, the rhythmic neural network 80, and the post-processing neural network 90 can use a common neural network model. A common neural network model (hereinafter referred to as composition neural network) is described.

3 shows an example of the structure of a neural network module. 3 shows a code neural network 30 as an example. The code neural network 30 includes a composition neural network 31 for composing a code and an evaluation neural network 32 for evaluating the output result of the composition neural network 31. [ Furthermore, the melody neural network 50, the rhythmic neural network 80, and the post-processing neural network 90 have the same configuration.

The training data uses a sequence of chords, melodies, rhythms, and the like of a song composed by humans. Furthermore, the training data may use a sequence of codes, melodies, rhythms, etc. of a song composed by artificial intelligence.

The composition neural network outputs a sequence of sequences from the learned data and delivers the output sequence to the evaluation neural network. Depending on the type of neural network, the sequence corresponds to a chord progression, a melody line, a rhythm, or a finished piece of music. When the evaluation neural network outputs the correct answer, it performs reinforcement learning in which the score of the composition neural network is deducted and the score is added when the wrong answer is outputted.

The composition neural network may include a learning unit for learning the sequence progression and an output unit for outputting the sequence progression. Composition The neural network can have the form of Auto Encoder. The auto encoder is a model that learns by using the unsupervised learning among the neural networks. An auto encoder consists of an encoder including an input layer and a hidden layer, and a decoder including a hidden layer and an output layer.

The composition neural network receives the feedback of the evaluation neural network and proceeds reinforcement learning while modifying the probability parameters. For example, the composition neural network modifies the probability parameter when the evaluation result of the evaluation neural network is correct, and reinforcement learning is performed by strengthening the parameter when the evaluation result is incorrect. In case of using auto encoder, it is possible to consider a method of learning by modifying the weight according to the output result of the evaluation neural network.

The evaluation neural network receives a randomly given sequence and determines whether it is a human-composed sequence or a sequence composed of artificial intelligence. The evaluation neural network adds the score when the discrimination result is correct, and subtracts the score when the discrimination result is incorrect. The evaluation neural network transmits the result of discrimination to the composition neural network.

The evaluation neural network uses the probabilistic parameters of the learning part and the learning part to learn labels that distinguish between homogeneous sequences and the source of the result, whether human or artificial intelligence, or whether the source of the input data is human composition or artificial intelligence And a discrimination unit for discriminating whether or not it is an artwork.

The evaluation neural network uses the sequence of the music composed by the human and the sequence composed by the composition artificial intelligence described above as the training data. The evaluation is performed while inputting a sequence composed of a human being in an arbitrary order and a sequence composed of artificial intelligence, and the internal probability parameter is adjusted so as to increase the accuracy of the evaluation result.

Once the evaluation neural network is constructed, the composition neural network receives a certain seed element and outputs a certain sequence, and the evaluation neural network evaluates the output sequence. As learning progresses, composing neural networks grow to output more artifacts of the same level as humans, so as to induce the wrong answers in the evaluation neural network, and the evaluation neural network grows in a direction to grasp it sharply.

Further, as shown in FIG. 2, the computer apparatus can use a separate neural network for codes, rhythms, and melodies. The computer device learns the artificial intelligence that selectively learns the subordinate musical structure Mars, rhythm, and melody, and outputs the sub structure, and can combine the results of these artificial intelligences.

The melody neural network 30 may be configured to receive and receive Mars and melody together so as to output a unified melody that does not deviate from harmony. You can also use separate neural networks for the baseline and melody lines. In this case, a separate baseline neural network can be used to generate the baseline. Baseline neural network construction is the same as melody neural network. The baseline neural network uses the baseline of existing music as learning data.

4 is another example showing a neural network model for music composition. 4 is an example of an artificial intelligence model for generating music. This is an example of a model different from Fig. 4 further includes seed neural network 10 and fitting neural network 20, unlike FIG.

The user can input the seed element in a simple manner as described in Fig. 1 without using professional information such as music score. What is needed for this is the seed neural network 10.

The seed neural network 10 receives a seed line from a user and converts it into a musical element. The fitting neural network 20 adjusts the melody to be output according to musical elements output from the seed neural network 10.

The user draws a line along a longitudinal or lateral line as shown in Fig. 1, or inputs a sound to input a seed line. The seed line means input information that a user inputs directly to a computer device or the like. For example, it is assumed that a line that is a graphical object is used as a seed line.

The user inputs a certain type of line through the interface device. The line can be matched to the change in pitch of the sound as the geometric information of change in elevation or change in left or right. The seed neural network 10 receives a line input by a user and outputs a musical motif. The musical motif is the aforementioned seed element. According to the artificial intelligence model, the seed neural network 10 outputs at least one of a seed code, a seed melody, and a seed rhythm. The seed neural network (10) is also prepared through learning in advance. The seed neural network 10 may be prepared using various types of machine learning or deep learning techniques. If you use the line graph element, you can use a neural network such as CNN (Convolutional Neural Network).

The fitting neural network 20 receives the seed element output by the seed neural network 10. [ For example, the fitting neural network 20 receives the seed melody and controls the melody to output a melody similar in pattern to the seed line information input by the user, compared with the melody output from the melody neural network 50. [ The fitting neural network 20 compares the result of repeating the seed melody or the seed melody with the melody of the specific section output by the melody neural network 50 and discards the melody output from the melody neural network 50 when the similarity degree of the pattern is out of the reference value , And may transmit a control command to the melody neural network 50 to generate a new melody line again. Through this observation, it is possible to extract a result corresponding to the seed information inputted by the user without violating the tendency of the music that the melody neural network 50 learned.

On the other hand, the information output by the post-processing neural network 90 may be a music file for playing music, or music information (text, score, etc.) including information such as melody / rhythm / harmony.

The present embodiment and drawings attached hereto are only a part of the technical idea included in the above-described technology, and it is easy for a person skilled in the art to easily understand the technical idea included in the description of the above- It will be appreciated that variations that may be deduced and specific embodiments are included within the scope of the foregoing description.

10: seed neural network
20: Fitting neural network
30: code neural network
31: Composition neural network
32: Evaluation Neural Network
50: melody neural network
80: rhythm neural network
90: Post-processing neural network
100: computer device

Claims (4)

Inputting a seed melody to a melody neural network in which a melody is learned, and extracting a certain melody line;
A step of the computer device inputting the rhythm of the rhythm into the rhythm neural network in which the rhythm is learned, and extracting a certain rhythm; And
The computer device comprising the step of adjusting the melody line in consideration of the rhythm according to the passage of time,
The melody neural network includes a melody learning neural network for learning melody generation by receiving sample data and an evaluation neural network for determining whether the output of the melody learning neural network is a sequence created by artificial intelligence or a sequence created by a human. Adjusting parameters of the learning neural network according to whether the determination result is correct,
The rhythmic neural network includes a rhythm learning neural network that receives sample data and learns rhythm generation and an evaluation neural network that determines whether the output of the rhythm learning neural network is a sequence created by artificial intelligence or a sequence created by a human. And adjusting the parameters of the rhythm learning neural network according to whether the result of the determination is correct or not. The method according to claim 1,
Further comprising the step of the computer device inputting a seed code into a code neural network where a code is learned and extracting a certain code progression, wherein the code neural network includes a code learning neural network that receives sample data and learns code generation, And an evaluation neural network for determining whether the output of the neural network is a sequence created by the artificial intelligence or a sequence created by a human being. In the code learning process, the parameter of the rhythm learning neural network is adjusted according to whether the determination result is correct,
Wherein the computer device further controls the melody line according to the code progression. The method according to claim 1,
Receiving a line or sound as a graphic element from a user; And
Further comprising the step of the computer device receiving the line or sound and inputting the line or sound into the seed neural network learned in advance to extract the seed melody and the seed rhythm,
Wherein the computer device inputs the extracted seed melody to the melody neural network and inputs the extracted seed rhythm to the rhythmic neural network. The method of claim 3,
Wherein when the similarity of the seed melody and the pattern is greater than or equal to a reference value, the computer device again inputs the seed melody to the melody neural network to start the melody line extraction again, Method of music composition using.

Images