Untitled

Onisawa Lab. Graduate School of System and Information Engineering, University of Tsukuba although a composition system needs to reflect user’s individual Abstract—This paper proposes a music/lyrics composing impressions, since the images to music or lyrics differ from person
system consisting of two sections, a lyric composing section and a
to person, the existing systems do not consider user’s music composing section, which reflects user's impressions of
theme of songs to music/lyrics. Love and nature are considered as
The purpose of this study is to reflect user’s impression and to theme in this paper. First of all, a user has a theme and image of
get the user’s satisfaction in composition of music/lyrics. Another lyrics to compose. The lyric composing section presents initial
purpose is that the system supports user’s composition of lyrics selected at random from database that is generated using
music/lyrics using user’s impression of music/lyrics even if a user existent lyrics and Markov Chain. If presented lyrics do not fit
has little knowledge of music or lyrics. user's image, a part of lyrics not fitting user's image is changed by
some other lyrics. When satisfied four lines lyrics are obtained,
In this article, Chapter 2 shows the structure of a music/lyrics the music composition section starts. This section composes music
composition system, Chapter 3 shows experiments using the fitting lyrics generated by the lyric composing section with the
presented system and discusses experimental results, and music composition system. The section presents combinations of
consideration, Chapter 4 shows the future work, and the final four lines lyrics and 16 measures music. A subject evaluates each
combination of lyrics and music whether they fit his/her image of
a song. According to subject's evaluation music melody is changed
by Genetic Algorithms and a part of lyrics are changed. These
procedures are repeated until satisfied combination of lyrics and
music melody is generated. In order to verify the validity of the
A. OUTLINE of MUSIC/LYRICS COMPOSITION SYSTEM presented system, subject experiments are performed.

First of all, a user has theme of song for composition of a song (music and lyrics), and keywords associated with a theme are considered. The outline of Music/Lyrics Composition System is shown in Fig.1. This system consists of two sections, a lyric OW,a lot of musical works in classical music, popular music, composing section and a music composing section. Four lines N odern music, screen music, computer music, and so on, are lyrics and a sixteen bars musical work is generated by this system. heard at restaurants or shopping centers, and music is a part of One-line lyrics correspond to four bars. our daily lives. Especially, Japanese popular song or nursery rhyme In the lyric composing section, lyrics are selected from is very familiar with us. Incidentally, Jean-Jacques Rousseau database constructed using Markov Chain [5]. When a user selects (1976) observed, “There was a sound, the soul grasped for it, and a theme, the system selects and shows to user the first candidate of there it had a ringing word” in “Essay on the origin of lyrics from lyrics database at random. If presented lyrics do not fit language”[1]. That is, music and words are inseparably related. user’s image, a part of lyrics not fitting user’s image is changed by Therefore, a song sometimes stirs a person’s heart than only words some other lyrics. If satisfied four lines lyrics are obtained, the or other music, which has only melodies. Moreover, a song has power of expression. There are many studies of automatic In the music composing section, music is generated with the composition music and automatic composition lyrics in order to music composition system [6], which uses Interactive Genetic realize music information processing and intelligent information Algorithms [7,8], a kind of genetic algorithms. In addition to processing of human. For instance, “LYRICA”[2] composes lyrics human evaluation and fitness functions, which evaluates whether automatically along with existing lyrics, and “VOCALOID”[3] is lyrics and melody’s rhythm fit each other. A user evaluates each known as a singing software. There is also a study on ”Music combination of lyrics and music whether they fit his/her image of a Composition by Interaction between Human and Computer”[4]. There are two ways to compose music/lyrics. One way is composing lyrics first, and the other is composing music first. However, human frequently composes music/lyrics while checking generated lyrics or music. It is important to adjust images of lyrics and the image of music. It is difficult to adjust their images, since lyrics and music are composed separately by them. Moreover, P is probability of elements b following elements a , ∑P = . n is the total number of morphemes, and m is the maximum number of b following a . If the number of b is less than m, all b have not elements. If b has no element for some i and j , its probability is assumed to be 0. If b has some element for j , b has the same element in some a . An example of words sets generated by Markov Chain is shown in Table1. Using the table, the following sentences are Let the first element of lyrics be c . First, select an element a at random. Here, a assumed to be selected and c Fig.1: Outline of Music/Lyrics Composition System is There. Element b following a is are or is. So, the next b at random. Here, are is assumed to 1) STRUCTURE of LYRIC COMPOSING SECTION be selected. Then, c is are. Next, select a of which element is The structure of the lyric composing section is shown in are. Here, a is are. Element Fig.2.When a user selects a theme, the section selects the first or blue. So, select elements again. By repeating this procedure, candidates of four lines lyrics from database of lyrics at random and shows to user. If presented lyrics do not fit user’s image, a part of lyrics not fitting user’s image is changed by some other lyrics. This section has two methods to change lyrics. The one is to change their considering lyrics context. The other is to choose lyrics at Table1: Example of words set generated by Markov Chain random. Furthermore, a user can also input lyrics composing by himself/herself. Show Four Lyrics(including Keyword) to User Fig.2: Structure of Lyric Composing Section 2) DATABASE of LYRIC COMPOSING SECTION The lyric composing section uses words sets [9] generated by using Markov Chain. The words sets are constructed by the analysis of existent lyrics using Markov Chain. First of all, existent lyrics are divided into morphemes by Morphological Analysis [11], In this section, three to four thousand lyrics are obtained by and are prepared the words sets {X shown as expression (1). the analysis of existent lyrics using Markov Chain. However, since these lyrics are obtained mechanically, although they are various types of lyrics, they don’t necessarily reflect user’s impressions of a theme. Therefore, keywords having relation to a theme are considered in order to select lyrics reflecting user’s impressions of a theme from 3 to 4 thousand lyrics. That is, lyrics including 2 K, n) in expression (1) consist of elements keywords are selected from words sets obtained by Markov Chain. b and probability P (i = ,.
Lyrics database is constructed by the collection of selected lyrics. morphemes divided, elements b are morphemes related to a , In order to consider the relation between lyrics this section also has database of lyrics considering connection of sentences as shown in Table 2. Table 2 shows that a sentence in condition i is presents a new melody fitting the number of words in new lyrics. liable to connect the one in condition i + 1. The database is used These procedures are repeated until the satisfied combination of when a user wants to change lyrics only one line. Table2: Database of lyrics considering connection of sentences Fifty candidates are arranged in high order 1) STRUCTURE of MUSIC COMPOSING SECTION The structure of the music composition section is shown in Fig.3. In this section, a music composition system [6] is used, which applies Interactive Genetic Algorithms. Human evaluation and the fitness function based on music theory are used in Interactive Genetic Algorithms. One melody with four measures is The numbers of words is assumed to be in the range from 10 considered as one phrase corresponding to one line lyric. to 25. Therefore, patterns of notes shown in Fig.5 are assumed to be used in this range. That is, 16th notes or 32nd notes are not use. This section generates melody lines to fit user’s image of a song by changing the order of notes at random. Fig.3: Structure of Music Composing Section In the application of genetic algorithms, one phrase corresponds to one individual, i.e., one chromosome. In addition to human evaluation and the fitness functions, which evaluates 3) FITTNESS whether lyrics and melody’s rhythm fit each other. This kind of In order to fit the melodies with lyrics, the difference between evaluation is described later on. The music composition section the number of notes and the number of words is evaluated. The generates music by the following procedures. The section presents fitness function is shown in expression (2). The fitness function combinations of four lines lyrics and 16 measures music with top value becomes low as the difference becomes large. 12 fitness evaluation values of the difference between the number of notes and that of words to a user, where the first 4 bars melody corresponds to the first one line lyrics, the next 4 bars to the second line lyrics, the next 4 bars to the third line one and the last four bars to the last line lyrics as shown in Fig.4. A user evaluates each combination of lyrics and music whether it fits his/her image of a song. According to user’s evaluation, music melody is changed by Genetic Algorithms. If lyrics do not fit the changed melody from M : The Number of Notes of Melodies user’s image point of view, a user modifies the lyrics using the lyric composing section. If the number of words in modified lyrics is different from that in old lyrics, the music composition section In order to verify the validity of the presented system, subject According to reference [8], the desirable number of search generations is in 10 to 20 generations for user’s tiredness in Interactive Generic Algorithms, the experiment for one subject is finished at the 10th generation. Subjects are twenty man/woman are all twenties. Nine subjects of them are inexperienced in composing music/lyrics, three are inexperienced in composing lyrics, five are inexperienced in composing music, and three of whom are experienced in composing music/lyrics. Each subject has a theme, love or nature. When a theme is love, keywords are making acquaintance, separating, etc. When theme is nature, keywords are mountain, sea, Fig.6: Interface of Lyric Composing Section flower, etc. The keywords are decided by the questionnaire done In the experiment subjects compose music/lyrics using interfaces shown in Fig.6 and Fig.7. Fig.6 shows the interface of the composition lyrics section, and Fig.7 shows the interface of the composition music section. After composing music/lyrics, subjects answer the following questionnaires. Questionnaire: a) Do you feel that the system helps you compose music/lyrics? b) Do you feel that the lyrics are poetic? c) Do you feel that the system helps you compose lyrics? d) Do you feel that the output lyrics fit your image? e) Do you feel that the system helps you compose melodies fitting f) Do you feel that the music fits your image? Each subject answers questionnaires with 5-points scale as follows: Fig.7: Interface of Music Composing Section B. EXPERIMENTAL RESULTS and CONSIDERATION For questionnaire d), comparison experiments are conducted between “database with keywords” and “database that without keywords”, in order to show that the usefulness of keywords for composition of music/lyrics reflecting user’s impression of a theme. Fig.8: Example of Output Theme is “Love” The result of Questionnaire c) is shown in Fig.11. The average evaluation point of subjects is 4.5 in 5-points scale. Only a subject, L3, don’t feel that the system helps user compose lyrics. Her mother tongue is not Japanese. Since the keywords are in Japanese, differences of images to lyrics among Japanese and non-Japanese seems to have an influence on the result. Fig.9: Result of Questionnaire a) The analysis result of Questionnaire a) is shown in Fig.9. The average evaluation point among all subjects is 4 in 5-points scale. It can be said almost all subjects feel that the system helps a user Fig.12: Result of Questionnaire d) It is found from the analysis result of Questionnaire a) that this system can help user composing music/lyrics. The analysis result of Questionnaire d) is shown in Fig.12. The average point among subjects is 3.6 in 5-points scale at the use of database with the keywords, and the average is 2.3 at the use of database without the keywords. According to t-test, there is significant difference between with keywords and without keywords at the significance level 1%. Lyrics generated by the section having database with keywords fit user’s image more than those by the section having database without keywords. Keywords fitting a theme of a song are important to generate lyrics fitting user’s image. However, some subjects give 2-points to the section having database with keywords. Therefore, other construction methods of database should be considered. It is found from the results of Questionnaires b) to d) that this section can help users compose lyrics reflecting user’s image. Fig.10: Result of Questionnaire b) The analysis result of Questionnaire b) is shown in Fig.10. The average evaluation point among subjects is 3.8 in 5-points scale. Three subjects don’t feel that generated lyrics are poetic. User L1, one of them, has comment the generated lyrics is a little unnatural grammatically. Fig.13: Result of Questionnaire e) The analysis result of Questionnaire e) is shown in Fig.13. The average among subjects is 3.8 in 5-points scale. Four subjects don’t feel that the system helps a user compose melodies fitting lyrics. User L8, one of them, has comment when good melodies are Fig.11: Result of Questionnaire c) generated, I want to compose a part of music by myself in order to This paper proposes a music/lyrics composing system with user’s impression of a theme. The system consists of two sections, the lyrics composing section and the music composing section. The former section has lyrics database obtained by the analysis of existing lyrics using Markov Chain. This section composes lyrics reflecting user’s impression of a theme by the use of lyrics database and keywords. The latter section applies Interactive Genetic Algorithms in which the difference between the number of words and that of notes is evaluated by prepared fitness function and a user evaluates whether melodies and lyrics fit user’s image. In order to verify the validity of the presented system, subject Fig.14: Result of Questionnaire f) experiments are performed. Each subject composes music/lyrics having a theme, love or nature using the presented system. After The analysis result of Questionnaire f) is shown in Fig.14. The the composition of music/lyrics, each subject answers average among subjects is 3.5 in 5-points scale. Four subjects don’t questionnaires from the following points of view: Whether feel that the generated music fits user’s image. User 5, one of them, combinations of composed music/lyrics fit subject’s image of a has comment it seems to be difficult to sing a song along the song or not and whether each subject can compose music/lyrics easily or not. The experimental results show that the presented From the above results of Questionnaires e) and f), this system can reflect user’s impression of melody, and can help user compose There are following problems as future works; reconsideration of the relation between melody and lyrics and the use of associate words database and/or synonym database. It is found from the results of Questionnaires a) to f) that this system can compose music/lyrics fitting user’s image and that this system can help inexperienced users compose music/lyrics. From the experimental results and user’s comments, the [1] Jean-Jacques Rousseau, Yoshihiko Kobayashi, “Essay on the origin of following two points are considered as future works. [2] Minako Yoshikawa, “Software Creation 2004”, 2004 (1)There are many comments such as this system often (sometimes) [3] YAMAHA Co., “VOCALOID”,http://www.vocaloid.com/jp/index.html composes lyrics that are hard to sing a song along generated [4] Shigeru Imai, Toshiharu Nagao, “Automation Composition Using melody. This is because the relation between melody and lyrics is Genetic Algorithm”, Electronic telecommunication society technology considered only by the difference between the number of words [5] Masashi Kowada, “Markov Chain”, Hakujitsusha, 1973 and that of notes as shown in expression (1). The relation between [6] M.Unehara, T.Onisawa, “Music Composition by Interaction between melody and lyrics should be also considered from other points of Human and Computer”, New Generation Computing vol.23, no.2, view. For example, it is important to consider the rhythm of set of words so that the present system can generate the melody fitting the [7] Hitoshi Iba, Japanese Society for Artificial Intelligence, “Theory of evolution calculation method”, Ohmsha, 2005. [8] Hiroaki Kitano, “Genetic Algorithm 4”, Sangyotosho, 2000 (2) The present system greatly depends on the lyrics database [9] Automatic Automatic generation of sentences with Markov chain, obtained by the analysis of existing lyrics using Markov Chain. http://kentarok.org/mt/2004/04/15/000330 Therefore, if the number of themes increases and various [10] Dear nursery rhyme, singing, and nursery rhyme, individual impressions should be considered, the system cannot http://www.mahoroba.ne.jp/~gonbe007/hog/shouka/00_songs.html Kenji Yoshimura, “Base of natural language processing”, compose lyrics reflecting user’s image. It is necessary to use associate words database and/or synonym database as well as the [12] Yuu Kitagawa, “Handbook of Music Theory”, Rittor Music, 1999

Source: http://isis2007.fuzzy.or.kr/submission/upload/A1211.pdf