Search

International audience; The general aim of this article is to sketch the processes of understanding music as a set of phenomena intricately involved with the IT novelties within the Slovenian research community from the three basic perspectives of ethno/musicologist, librarian (essential when approaching music computationally), and IT. It assesses what these perspectives have brought to the understanding of music, but more importantly indicates thatthe three perspectives are not arbitrary. Together they form digital musicology in Slovenia. First, it lays out the Slovenian contributions to e-music research, and illustrates academic specificities, both Slovenian and global. Lastly, the article offers some thoughts on current academic challenges and reflects on possible solutions.

In recent years, deep architectures, most commonly based on neural networks, have advanced the state of the art in many research areas. Due to the popularity and the success of deep neural-networks, other deep architectures, including compositional models, have been put aside from mainstream research. This dissertation presents the compositional hierarchical model as a novel deep architecture for music processing. Our main motivation was to develop and explore an alternative non-neural deep architecture for music processing which would be transparent, meaning that the encoded knowledge would be interpretable, trained in an unsupervised manner and on small datasets, and useful as a feature extractor for classification tasks, as well as a transparent model for unsupervised pattern discovery. We base our work on compositional models, as compositionality is inherent in music. The proposed compositional hierarchical model learns a multi-layer hierarchical representation of the analyzed music signals in an unsupervised manner. It provides transparent insights into the learned concepts and their structure. It can be used as a feature extractor---its output can be used for classification tasks using existing machine learning techniques. Moreover, the model's transparency enables an interpretation of the learned concepts, so the model can be used for analysis (exploration of the learned hierarchy) or discovery-oriented (inferring the hierarchy) tasks, which is difficult with most neural network based architectures. The proposed model uses relative coding of the learned concepts, which eliminates the need for large annotated training datasets that are essential in deep architectures with a large number of parameters. Relative coding contributes to slim models, which are fast to execute and have low memory requirements. The model also incorporates several biologically-inspired mechanisms that are modeled according to the mechanisms that exists at the lower levels of human perception (eg~ lateral inhibition in the human ear) and that significantly affect perception. The proposed model is evaluated on several music information retrieval tasks and its results are compared to the current state of the art. The dissertation is structured as follows. In the first chapter we present the motivation for the development of the new model. In the second chapter we elaborate on the related work in music information retrieval and review other compositional and transparent models. Chapter three introduces a thorough description of the proposed model. The model structure, its learning and inference methods are explained, as well as the incorporated biologically-inspired mechanisms. The model is then applied to several different music domains, which are divided according to the type of input data. In this we follow the timeline of the development and the implementation of the model. In chapter four, we present the model's application to audio recordings, specifically for two tasks: automatic chord estimation and multiple fundamental frequency estimation. In chapter five, we present the model's application to symbolic music representations. We concentrate on pattern discovery, emphasizing the model's ability to tackle such problems. We also evaluate the model as a feature generator for tune family classification. Finally, in chapter six, we show the latest progress in developing the model for representing rhythm and show that it exhibits a high degree of robustness in extracting high-level rhythmic structures from music signals. We conclude the dissertation by summarizing our work and the results, elaborating on forthcoming work in the development of the model and its future applications. S porastom globokih arhitektur, ki temeljijo na nevronskih mrežah, so se v zadnjem času bistveno izboljšali rezultati pri reševanju problemov na več področjih. Zaradi popularnosti in uspešnosti teh globokih pristopov, temelječih na nevronskih mrežah, so bili drugi, predvsem kompozicionalni pristopi, odmaknjeni od središča pozornosti raziskav. V pričujoči disertaciji se posvečamo vprašanju, ali je mogoče razviti globoko arhitekturo, ki bo presegla obstoječe probleme globokih arhitektur. S tem namenom se vračamo h kompozicionalnim modelom in predstavimo kompozicionalni hierarhični model kot alternativno globoko arhitekturo, ki bo imela naslednje značilnosti: transparentnost, ki omogoča enostavno razlago naučenih konceptov, nenadzorovano učenje in zmožnost učenja na majhnih podatkovnih bazah, uporabnost modela kot izluščevalca značilk, kot tudi zmožnost uporabe transparentnosti modela za odkrivanje vzorcev. Naše delo temelji na kompozicionalnih modelih, ki so v glasbi intuitivni. Predlagani kompozicionalni hierarhični model je zmožen nenadzorovanega učenja večnivojske predstavitve glasbenega vhoda. Model omogoča pregled naučenih konceptov skozi transparentne strukture. Lahko ga uporabimo kot generator značilk -- izhod modela lahko uporabimo za klasifikacijo z drugimi pristopi strojnega učenja. Hkrati pa lahko transparentnost predlaganega modela uporabimo za analizo (raziskovanje naučene hierarhije) pri odkrivanju vzorcev, kar je težko izvedljivo z ostalimi pristopi, ki temeljijo na nevronskih mrežah. Relativno kodiranje konceptov v samem modelu pripomore k precej manjšim modelom in posledično zmanjšuje potrebo po velikih podatkovnih zbirkah, potrebnih za učenje modela. Z vpeljavo biološko navdahnjenih mehanizmov želimo model še bolj približati človeškemu načinu zaznave. Za nekatere mehanizme, na primer inhibicijo, vemo, da so v človeški percepciji prisotni na nižjih nivojih v ušesu in bistveno vplivajo na način zaznave. V modelu uvedemo prve korake k takšnemu načinu procesiranja proti končnemu cilju izdelave modela, ki popolnoma odraža človeško percepcijo. V prvem poglavju disertacije predstavimo motivacijo za razvoj novega modela. V drugem poglavju se posvetimo dosedanjim objavljenim dosežkom na tem področju. V nadaljnjih poglavjih se osredotočimo na sam model. Sprva opišemo teoretično zasnovo modela in način učenja ter delovanje biološko-navdahnjenih mehanizmov. V naslednjem koraku model apliciramo na več različnih glasbenih domen, ki so razdeljene glede na tip vhodnih podatkov. Pri tem sledimo časovnici razvoja in implementacijam modela tekom doktorskega študija. Najprej predstavimo aplikacijo modela za časovno-frekvenčne signale, na katerem model preizkusimo za dve opravili: avtomatsko ocenjevanje harmonij in avtomatsko transkripcijo osnovnih frekvenc. V petem poglavju predstavimo drug način aplikacije modela, tokrat na simbolne vhodne podatke, ki predstavljajo glasbeni zapis. Pri tem pristopu se osredotočamo na odkrivanje vzorcev, s čimer poudarimo zmožnost modela za reševanje tovrstnih problemov, ki je ostalim pristopom še nedosegljivo. Model prav tako evalviramo v vlogi generatorja značilk. Pri tem ga evalviramo na problemu melodične podobnosti pesmi in razvrščanja v variantne tipe. Nazadnje, v šestem poglavju, pokažemo zadnji dosežek razvoja modela, ki ga apliciramo na problem razumevanja ritma v glasbi. Prilagojeni model analiziramo in pokažemo njegovo zmožnost učenja različnih ritmičnih oblik in visoko stopnjo robustnosti pri izluščevanju visokonivojskih struktur v ritmu. V zaključkih disertacije povzamemo vloženo delo in rezultate ter nakažemo nadaljnje korake za razvoj modela v prihodnosti.

In recent years, deep architectures, most commonly based on neural networks, have advanced the state of the art in many research areas. Due to the popularity and the success of deep neural-networks, other deep architectures, including compositional models, have been put aside from mainstream research. This dissertation presents the compositional hierarchical model as a novel deep architecture for music processing. Our main motivation was to develop and explore an alternative non-neural deep architecture for music processing which would be transparent, meaning that the encoded knowledge would be interpretable, trained in an unsupervised manner and on small datasets, and useful as a feature extractor for classification tasks, as well as a transparent model for unsupervised pattern discovery. We base our work on compositional models, as compositionality is inherent in music. The proposed compositional hierarchical model learns a multi-layer hierarchical representation of the analyzed music signals in an unsupervised manner. It provides transparent insights into the learned concepts and their structure. It can be used as a feature extractor—its output can be used for classification tasks using existing machine learning techniques. Moreover, the model’s transparency enables an interpretation of the learned concepts, so the model can be used for analysis (exploration of the learned hierarchy) or discovery-oriented (inferring the hierarchy) tasks, which is difficult with most neural network based architectures. The proposed model uses relative coding of the learned concepts, which eliminates the need for large annotated training datasets that are essential in deep architectures with a large number of parameters. Relative coding contributes to slim models, which are fast to execute and have low memory requirements. The model also incorporates several biologically-inspired mechanisms that are modeled according to the mechanisms that exists at the lower levels of human perception (e.g. lateral inhibition in the human ear) and that significantly affect perception. The proposed model is evaluated on several music information retrieval tasks and its results are compared to the current state of the art. The dissertation is structured as follows. In the first chapter we present the motivation for the development of the new model. In the second chapter we elaborate on the related work in music information retrieval and review other compositional and transparent models. Chapter three introduces a thorough description of the proposed model. The model structure, its learning and inference methods are explained, as well as the incorporated biologically-inspired mechanisms. The model is then applied to several different music domains, which are divided according to the type of input data. In this we follow the timeline of the development and the implementation of the model. In chapter four, we present the model’s application to audio recordings, specifically for two tasks: automatic chord estimation and multiple fundamental frequency estimation. In chapter five, we present the model’s application to symbolic music representations. We concentrate on pattern discovery, emphasizing the model’s ability to tackle such problems. We also evaluate the model as a feature generator for tune family classification. Finally, in chapter six, we show the latest progress in developing the model for representing rhythm and show that it exhibits a high degree of robustness in extracting high-level rhythmic structures from music signals. We conclude the dissertation by summarizing our work and the results, elaborating on forthcoming work in the development of the model and its future applications.

Although masses of electronic learning materials are being regularly created for e-learning purposes, there is still difficult for a teacher to find a suitable material for a particular teaching situation. Instead of adapting lessons to suit the available learning content, the teachers should actively adjust the learning content itself to make it suit their needs for use in the class. Despite having good ideas, not all teachers are capable of creating an attractive learning content, or even just customising it due to the lack of programming knowledge and inadequate ICT-usage skills. Our goal is to bring together two distinct communities, teachers and programmers, to work together, share ideas, and brainstorm, with the common goal to benefit from this mashing in providing useful materials for enhancing learning experience. The portal is aimed as a meeting place for teachers with innovative ideas for new e-content and technology buffs wishing to contribute their knowledge to common good.

The thesis offers an implementation of a new approach for automatic modelling and music information retrieval. The model is based on a visual compositional hierarchical model, which provides good results in the area of categorization of visual objects in pictures. We translate the approach to the music domain by implementing and visualizing the model. The approach uses knowledge of psychoacoustic music modelling for the preparation of input audio clips. The model is tested on a set of four basic piano chords, played through all eight octaves. We achieve 83% classification accuracy by classifying into major and minor chords and 78% classification accuracy by classifying into major, minor, diminished and augmented chords. The model is also tested on a set of songs composed by The Beatles. We successfully classify 60% of chords into major and minor chord classes. We find the translation of the compositional hierarchical model to the music domain successful and worth of further development.

[TODO] Add abstract here.

Diplomska naloga rešuje problematiko nedostopnosti ustrezne programske rešitve za učence osnovnih glasbenih šol. Otrokom moramo zagotoviti rešitev v slovenskem jeziku, ki jo lahko igrajo na kateri koli digitalni napravi. Tako rešitev naredimo dostopno vsakemu otroku ter jo prilagodimo zahtevam mlajših otrok. Raziskali smo metodologije evalvacije, s katerimi lahko dobimo merljive rezultate pri izdelavi iteracij. Identificirali smo različne fokusne skupine, nakar smo lahko nadaljevali na evalvacijo izdelka, ki smo ga že izdelali v več iteracijah. S pomočjo metodologij smo nato oblikovali izdelek, ki smo ga preizkušali z otroci, ki obiskujejo osnovno glasbeno šolo. Evalvacija oziroma intervjuji so pritrdili ustreznost ter razumevanje vmesnika. Otroci ter tudi učitelji so podali veliko odzivov, ki bodo pomagali pri naslednjih različicah rešitve. Končni izdelek naloge je iterativno grajen vmesnik, prilagojen zahtevam in potrebam fokusnih skupin. The thesis solves the problem of the lack of a suitable software solution for primary music school students. We need to provide children with a solution in Slovenian that they can play on any digital device. This way we make the solution accessible to every child and adapt it to the requirements of younger children. We have explored evaluation methodologies that can produce measurable results when iterating. We identified different focus groups, and then we were able to move on to evaluating a product that we had already produced in several iterations. We then used the methodologies to design a product that we tested with children attending a primary music school. The evaluation or interviews confirmed the relevance and understanding of the interface. The children as well as the teachers gave a lot of feedback which will help in the next versions of the solution. The final product of the exercise is an iteratively built interface, tailored to the requirements and needs of the focus groups.

Otrok se lahko ritmičnega posluha nauči le v glasbeni šoli. Otroci, ki so pri pouku prisotni redkeje, so deležni manj vaje in posledično dosežejo slabše rezultate pri preizkusih znanja. Otroci ritmične vaje težko samostojno vadijo doma. Na trgu sicer obstajajo pripomočki za urjenje ritmičnega posluha, vendar niso skladni s slovenskim učnim načrtom. Te pripomočke smo raziskali in ugotavljali, kateri tipi vaj so najprimernejši in kateri grafični vmesniki omogočajo najboljšo uporabniško izkušnjo. Sistem za generiranje ritmičnih vaj smo razširili v okviru platforme Trubadur in razvili tri igre za urjenje ritmičnega posluha: igro ritmičnega nareka, igro ritmičnega kviza in igro igranja ritma. Sistem smo testirali na skupini otrok, ki obiskujejo osnovno glasbeno šolo oziroma se nameravajo vanjo vpisati. Rezultati testiranja so pokazali, da sta igra ritmičnega nareka in igra ritmičnega kviza bolj ustrezni od igre igranja ritma. Otroci so bili s sistemom na splošno zadovoljni in so podali veliko koristnih povratnih informacij za nadaljnji razvoj. A child can only develop a good ear for rhythm in a music school. Children who attend lessons less often receive less practice and, as a result, achieve worse results in knowledge tests. It is difficult for children to practice rhythmic exercises independently at home. There are applications on the market for training rhythmic listening, but they are not compatible with the Slovenian curriculum. We researched these tools and determined which types of exercises are most suitable and which graphical interfaces provide the best user experience. We expanded the system for generating rhythmic exercises within the framework of the Troubadour platform and developed three games for training rhythmic listening: a rhythmic dictation game, a rhythmic quiz game, and a rhythm playing game. We tested the system on a group of children attending or planning to enroll in a primary music school. The test results showed that the rhythmic dictation game and the rhythmic quiz game are more relevant than the rhythm playing game. The children were generally satisfied with the system and provided a lot of useful feedback for further development.

Namen diplomskega dela je razvoj sistema za avdio podpisovanje in razpoznavo posnetkov glasbe. Avdio podpisovanje se uporablja za primerjavo zvočnih posnetkov in je bolj ali manj odporno na šum ter druge motnje iz okolja. V okviru diplomskega dela smo raziskali obstoječe sisteme za avdio podpisovanje, podrobno opisali razviti sistem ter ga evalvirali. Z njim smo indeksirali 900 glasbenih del (54 ur posnetkov) in izmerili njegovo točnost razpoznave glasbe, hitrost indeksiranja, hitrost razpoznave posnetkov ter prostorsko zahtevnost. Sistem je dovolj hiter in točen za uporabo v manjših aplikacijah, kot sta deduplikacija arhivskih posnetkov in razpoznava glasbe, predvajane po radiu. Sistem pričakovano dosega slabše rezultate pri razpoznavi glasbe, posnete v živo, zato za tovrstno uporabo ni primeren. This thesis aims to develop a system for audio fingerprinting and recognition of recorded music. Audio fingerprinting is a process used to compare audio recordings regardless of noise or other environmental factors. Throughout this work, we explore several audio fingerprinting methods and provide a detailed description of the developed system. The system was used to index 900 songs (54 hours of recorded music) and evaluate its accuracy, indexing speed, query speed, and spatial consumption. The system is found to be sufficiently fast and accurate for use in smaller applications like archive deduplication and recognition of songs played over the radio. However, it lacks sufficient accuracy to be used to identify music played in noisy environments.