BDB Bertsolaritzaren datu-basea

[EU] Ikerketa lan honetan bi gai nagusi landu dira: musikaren sorkuntza automatikoa eta sailkapena. Musikaren sorkuntzarako bertso doinuen corpus bat hartu da abiapuntu moduan doinu ulergarri berriak sortzeko gai den metodo bat sortzeko. Doinuei ulergarritasuna hauen barnean dauden errepikapen egiturek ematen dietela suposatu da, eta metodoaren hiru bertsio nagusi aurkeztu dira, bakoitzean errepikapen horien definizio ezberdin bat erabiliz.Musikaren sailkapen automatikoan hiru ataza garatu dira: generoen sailkapena, familia melodikoen taldekatzea eta konposatzaileen identifikazioa. Musikaren errepresentazio ezberdinak erabili dira ataza bakoitzerako, eta ikasketa automatikoko hainbat teknika ere probatu dira, emaitzarik hoberenak zeinek ematen dituen aztertzeko.Gainbegiratutako sailkapenaren alorrean ere binakako sailkapenaren gainean lana egin da, aurretik existitzen zen metodo bat optimizatuz. Hainbat datu baseren gainean probatu da garatutako teknika, baita konposatzaile klasikoen piezen ezaugarriez osatutako datu base batean ere.

[EN] In the recent years, with the advances on computation sciences, many works on different topics related to music have been developed in order to make easier tasks that had to be made by hand. Different approaches have been developed, working both with audio and symbolic data, and several techniques have also been used, such as machine learning, rule based or deep learning methods. The applications to these methods are also diverse, and they go from automatic music generation to music recommendation systems. In this research work two main topics have been studied: automatic music generation and classification. For the automatic music generation part a corpus of bertso melodies has been used to create a method that is able to generate new understandable tunes. The works made in this field have been based on the idea that having structures of repeated segments within a piece, at least in some abstract level, is necessary to understand it when we listen to it. Three versions of the generation method are presented in this manuscript, using different repetition structure definitions. In the music classification part three main tasks have been tackled: genre classification, tune family identification and composer recognition. Different music representations have been used in the different tasks. Several machine learning techniques have also been applied in order to test which ones offer better classification results. Finally, work on supervised classification has also been made, specifically on class binarization, where an attempt to optimize a previous binarization technique has been made. This optimization has been applied to several databases, among them one that consists of features of musical pieces of well known classical composers.

1 Introduction ...1
1.1 Motivation ...2
1.2 Music representation ...2
1.2.1 MIDI ...2
1.2.2 Viewpoint representation ...4
1.3 Contributions ...5
1.3.1 Automatic Music Generation ...5
1.3.2 Supervised Classification ...6
1.3.3 Automatic Music Classification ...6
1.4 Thesis structure ...7
2 Automatic music generation ...9
2.1 Introduction ...9
2.2 State of the art ...11
2.3 Contributions ...13
2.3.1 Corpus ...14
2.3.2 Note level coherence ...15
2.3.3 Music generation with a coherence structure with multiple
abstraction levels ...16
2.3.4 Rhythmic structure in template ...27
2.4 Conclusions and future work ...32
3 Supervised Classification ...35
3.1 Introduction ...35
3.2 Dynamic Classifier Selection ...37
3.3 Class binarization ...38
3.3.1 OVO...39
3.3.2 Dynamic Classifier Selection and OVO ...40
3.4 Contributions ...41
3.4.1 PSEUDOVO ...41
3.5 Conclusions ...46
4 Music Classification ...49
4.1 Introduction ...49
4.2 State of the art ...51
4.3 Contributions ...53
4.3.1 Unsupervised classification of bertso melodies ...53
4.3.2 Tune family classification with pattern covering ...58
4.3.3 Composer recognition with matrix representation ... 63
4.4 Conclusions ...66
5 Conclusions and future work ...69
I Articles Related to Automatic Music Generation... 71
6 Transformation of a bertso melody with coherence ...73
7 Melody Transformation with Semiotic Patterns ...79
8 Statistics based music generation approach considering both rhythm and melody coherence ...93
II Article Related to Supervised Classification ...113
9 Problems Selection Under Dynamic selection of the best base classifier
in One versus One: PSEUDOVO ...115
III Articles Related to Music Classification ...143
10 Towards the use of similarity distances to music genre classification: A comparative study ...145
11 Melody classification with pattern covering ...165
12 On the Use of Matrix Based Representation to Deal with Automatic
Composer Recognition ...173
Bibliography ...181