Automatický přepis hudby a odhad vícero znějících tónu jsou stále výzvou v oblasti dolování informací z hudby. Moderní systémy jsou založeny na různých technikách strojového učení pro dosažení co nejpřesnějšího přepisu hudby. Některé z nich jsou také omezeny na konkrétní hudební nástroj nebo hudební žánr, aby se snížila rozmanitost analyzovaného zvuku. V této práci je navrženo, vyhodnoceno a porovnáváno několik systémů pro konverzi nahrávek elektrické kytary  do MIDI souború, založených na různých technikách strojového učení a technikách spektrální analýzy.

This diploma work deals with music genre recognition using the techniques of Music Information Retrieval. It contains a brief description of the principle of this research area and its subfield called Music Genre Recognition. The following chapter includes selection of the most suitable parameters for describing music genres. This work further characterizes machine learning methods used in this field of research. The next chapter deals with the descriptions of music datasets created for genre classification studies. Subsequently, there is a draft and evaluation of the system for music genre recognition. The last part of this work describes the results of partial parameter analysis, dependence of genre classification accuracy on the amount of parameters and contains a discussion on the causes of classification accurancy for the individual genres.

This master thesis deals with systems for detecting rhythmic structures of music recordings. The field of Music Information Retrieval (MIR) allows us to examine the harmonic and tonal properties of music, rhythm, tempo, etc., and has uses in academic and commercial sphere. Various algorithms are used in the detection of rhythmic structures. However, today, most new methods use neural networks. This work aims to summarize the current research results of systems for detecting beats and tempo, to describe methods of calculating and evaluating the parameters of music recordings, and to implement a program that allows comparison of available detection systems. The result of the work is a script in the Python language, which uses six different systems to detect the rhythmic structure of test recordings. It then checks the outputs of the algorithms according to the given reference and compares the given systems with each other using several evaluation values. It uses two datasets as a reference—one of them is publicly available and the other was created by the author of this thesis (including annotations, i.e., reference beat times, for the sample recordings). The program allows user to see the results in graphs and play any of the sample recordings with detected beat times.

La musique a la capacité d’induire et moduler les émotions, décomposées en deux dimensions : le niveau d’activation (relaxant-stimulant) et la valence émotionnelle (déplaisant-plaisant). Une façon de mesurer objectivement la valence musicale est par le réflexe acoustique de sursaut, une réaction de défense qui consiste en un clignement de l’oeil provoqué par un bruit fort et inattendu. Le réflexe est renforcé par la musique déplaisante et inhibée par la musique plaisante. Cependant, l’effet du niveau d’activation émotionnelle lors de l’écoute musicale demeure inconnu. Cette étude a donc pour objectif d’examiner la modulation du réflexe acoustique de sursaut par la musique stimulante et relaxante jugée plaisante. Basée sur les résultats d’études antérieures avec des images, notre hypothèse était que le réflexe serait plus faible dans la condition stimulante que dans la condition relaxante. Dans un devis intrasujet, 47 participants ont écouté de la musique relaxante et stimulante. Des bruits blancs courts et forts ont été rajoutés par-dessus les extraits afin de provoquer le réflexe de sursaut, dont son amplitude et sa latence ont été mesurées par électromyographie. Les résultats ont ensuite été comparés à ceux d’une condition non-musicale, constituée de sons environnementaux plaisants, afin d’explorer si la musique est plus efficace pour inhiber le réflexe. Finalement, des caractéristiques acoustiques, telles que la clarté de la pulsation, la densité acoustique, la dissonance et l’énergie, ont été extraites puis comparées entre les trois conditions pour explorer leur relation avec les paramètres du réflexe. Les résultats rapportent une modulation de la latence du réflexe de sursaut, dans laquelle celle-ci est plus longue dans la condition stimulante comparée à la condition relaxante. Cependant, aucune différence au niveau de l’amplitude n’a été observée. Seule la latence serait donc sensible au niveau d’activation des émotions musicales lorsque la musique est plaisante. Ensuite, la latence dans la condition non-musicale était aussi longue que celle dans la condition stimulante, suggérant que la musique n’est pas plus efficace que les sons non-musicaux pour inhiber le réflexe de sursaut. Finalement, comme l’amplitude et la latence n’ont pas le même patron de réponses, cette étude suggère que le réflexe de sursaut est aussi modulé par le traitement des caractéristiques acoustiques et que ceux-ci ont un effet différent sur ces deux paramètres. En conclusion, la latence du réflexe acoustique de sursaut est une bonne méthode pour mesurer le niveau d’activation des émotions musicales. De futures recherches pourront utiliser le paradigme de la modulation affective du réflexe de sursaut pour mesurer les effets des émotions musicales selon des facteurs individuels tels que l’âge et la dépression. / Music has the capacity to evoke and modulate emotions, divided by two dimensions: arousal (relaxing-stimulating), and valence (unpleasant-pleasant). Musical valence can be objectively measured by the acoustic startle reflex, a defensive reaction consisting of an eye blink provoked by a short and loud noise. This reflex is facilitated by unpleasant music and inhibited by pleasant music. However, the arousal effect while listening to music on the startle reflex remains unknown. This study therefore aims to explore the affective startle modulation by stimulating and relaxing music. In a within-subjects design, 47 participants listened to stimulating music, relaxing music and non-musical sounds. White noises (50 ms, 105 dB(A)) were added over the excerpts to induce startle while eyeblink magnitude and latency were measured by electromyography. Excerpts’ acoustic features were then extracted and compared through experimental conditions to explore their effect on startle modulation. Startle latency was longer in the stimulating condition compared to the relaxing one, but no differences in magnitude were found, partially confirming our predictions. Exploratory analyses suggest that startle modulation is also attributed to bottom-up processes of acoustic features, and that these latter impact differently magnitude and latency. In conclusion, this study highlights startle latency measure efficiently emotional arousal while listening to music, allowing future research to use the paradigm of affective startle reflex modulation to evaluate the effect of music on emotions considering individual factors, such as age and depression. It also paves the way for comparisons of the effect of emotions and acoustic features processes on the startle reflex modulation.

Automatic Music Transcriptions systems such as ScoreCloud aims to convert audio signals to sheet music. The information contained in sheet music can be divided into increasingly descriptive layers, where most research on Automatic Music Transcription is restricted on note-level transcription and disregard expressive markings such as legato slurs. In case of violin playing, legato can be determined from the articulated, "hard" onsets that occur on the first note of a legato slur. We detect hard onsets in violin recordings by three different methods — two based on signal processing and one on Convolutional Neural Networks. ScoreCloud notes are then labeled as articulated or slurred, depending on the distance to the closest hard onset. Finally, we construct legato slurs between articulated notes, and count the number of notes where the detected slur label matches ground-truth. Our best-performing method correctly labels notes in 82.9% of the cases, when averaging on the test set recordings. The designed system serves as a proof-of-concept for including expressive notation within Automatic Music Transcription. Vibrato was seen to have a major negative impact on the performance, while the method is less affected by varying sound quality and polyphony. Our system could be further improved by using phase input, data augmentation, or high-dimensional articulation representations. / System för automatisk musiktranskription såsom ScoreCloud syftar till att konvertera ljudsignaler till notskrift. Informationen i en notbild kan delas in i flera lager med en ökande nivå av beskrivning, där huvuddelen av forskningen har begränsats till transkription av noter och har bortsett från uttrycksmarkeringar såsom legatobågar. I fallet med violin kan legato bestämmas från de artikulerade, ’hårda’ ansatser som uppkommer vid den första noten i en legatobåge. Vi detekterar här hårda ansatser i inspelningar av violin genom tre olika metoder — två baserade på signalbehandling och en baserat på faltningsnätverk. Noter från ScoreCloud märks sedan som artikulerade eller bundna, beroende på det närmaste avståndet till en hård ansats. Slutligen konstrueras legatobågar mellan artikulerade noter, och vi räknar antalet noter där den predicerade märkningen stämmer med den sanna. Vår bäst presterande metod gör en korrekt märkning i 82.9% i genomsnitt taget över testinspelningarna. Vårt system validerar idén att innefatta uttrycksmarkeringar i automatisk musiktranskription. Vibrato observerades påverka resultatet mycket negativt, medan metoden är mindre påverkad av varierande ljudkvalitet och polyfoni. Vårt system kan förbättras ytterligare genom användandet av fas i indata, datautvidgning och högdimensionella representationer av artikulation.

Thanks to the advancement of technological progress in networking and signal processing, we can access a large amount of musical content. In order for users to search among these vast catalogs, they need to have access to music-related information beyond the pure digital music file. Manual annotation of music is too expensive, therefore automated annotation would be of great use. A meaningful description of the musical pieces requires the incorporation of information about the instruments present in them. In this work, we present an approach for musical instrument recognition using the scattering transform, which is a transformation that gives a translation invariant representation, that is stable to deformations and preserves high frequency information for classication. We study recognition in both singleinstrument and multiple-instrument contexts. We compare the performance of models using the scattering transform to those using other standard features. We also examine the impact of the amount of training data. The experiments carried out do not show a clear superior performance of either feature representation. Still, the scattering transform is worth taking into account when choosing a way to extract features if we want to be able to characterize non-stationary signal structures. / Tack vare den tekniska utvecklingen i nätverk och signalbehandling kan vi få tillgång till en stor mängd musikaliskt innehåll. For att användare ska söka bland dessa stora kataloger måste de ha tillgång till musikrelaterad information utöver den rena digitala musikfilen. Eftersom den manuella annotationsprocessen skulle vara för dyr måste den automatiseras. En meningsfull beskrivning av musikstyckena kräver införlivande av information om instrumenten som finns i dem. I det här arbetet presenterar vi en metod for igenkänning av musikinstrument med hjälp av den scattering transform, som är en transformation som ger en översattnings-invariant representation, som är stabil för deformationer och bevarar högfrekvensinformation för klassicering. Vi studerar igenkännande i både enskilda instrument- och flera instrumentförhållanden. Vi jämför modellerna med den scattering transforms prestanda med de som använder andra standardfunktioner. Vi undersöker också effekterna av mangden traningsdata. Experimenten som utförs visar inte en tydlig överlagsen prestanda for någon av representationsföreställningarna jämfört med den andra. Fortfarande är den scattering transform värd att ta hänsyn till när man väljer ett sätt att extrahera funktioner om vi vill kunna karakterisera icke-stationära signalstrukturer.

Perceptual features are qualitative features used to describe music properties in relation to human perception instead of typical musical theory concepts such as pitches and chords. This report describes a music discovery platform which uses three different methods of music playlist generation to investigate if and how perceptual features work when used for music discovery. One method abstracts away the complexity of perceptual features and the other two lets users use them directly. Two user testing sessions were performed to evaluate the browser and compare the different methods. Test participants found the playlist generation to work well in general, and especially found the method which uses emotions as an interface to be intuitive, enjoyable and something they would use to find new music. The other two methods which let users directly interact with perceptual features were less popular, especially among users without musical education. Overall, using perceptual features for music discovery was successful, although methods should be chosen with the intended audience in mind. / Perceptuella särdrag är kvalitativt framtagna särdrag som beskriver musik med fokus på mänsklig perception snarare än musikteoribegrepp som tonhöjd och ackord. Den här rapporten beskriver en musikhemsida som använder tre olika metoder för att generera spellistor med avsikt att undersöka om och hur perceptuella särdrag fungerar för att hitta ny musik. En metod abstraherar bort perceptuella särdragens komplexitet och de andra två metoderna låter testare använda dem utan abstraktion. Två användbarhetstest utfördes för att utvärdera musikhemsidan och jämföra de olika metoderna. Testanvändare tyckte överlag att genereringen av spellistor fungerade bra och att speciellt metoden som använde känslor som gränssnitt var intuitiv, rolig att använda och en metod de skulle använda för att hitta ny musik. De andra två metoderna som tillät användare att direkt använda perceptuella särdrag var mindre populära, speciellt bland användare utan musikutbildning. Överlag var användandet av perceptuella särdrag för att hitta musik en framgång, dock bör metoderna väljas utifrån användarnas kunskap.

Der Beitrag beschreibt ein laufendes Projekt zur computergestützten Erschließung und Analyse einer großen Sammlung handschriftlicher Liedblätter mit Volksliedern aus dem deutschsprachigen Raum. Am Beispiel dieses praktischen Projekts werden Chancen und Herausforderungen diskutiert, die der Einsatz von Digital Humanities-Methoden für den Bereich der Musikwissenschaft mit sich bringt.

L'apprentissage machine (AM) est un outil important dans le domaine de la recherche d'information musicale (Music Information Retrieval ou MIR). De nombreuses tâches de MIR peuvent être résolues en entraînant un classifieur sur un ensemble de caractéristiques. Pour les tâches de MIR se basant sur l'audio musical, il est possible d'extraire de l'audio les caractéristiques pertinentes à l'aide de méthodes traitement de signal. Toutefois, certains aspects musicaux sont difficiles à extraire à l'aide de simples heuristiques. Afin d'obtenir des caractéristiques plus riches, il est possible d'utiliser l'AM pour apprendre une représentation musicale à partir de l'audio. Ces caractéristiques apprises permettent souvent d'améliorer la performance sur une tâche de MIR donnée. Afin d'apprendre des représentations musicales intéressantes, il est important de considérer les aspects particuliers à l'audio musical dans la conception des modèles d'apprentissage. Vu la structure temporelle et spectrale de l'audio musical, les représentations profondes et multiéchelles sont particulièrement bien conçues pour représenter la musique. Cette thèse porte sur l'apprentissage de représentations de l'audio musical. Des modèles profonds et multiéchelles améliorant l'état de l'art pour des tâches telles que la reconnaissance d'instrument, la reconnaissance de genre et l'étiquetage automatique y sont présentés. / Machine learning (ML) is an important tool in the field of music information retrieval (MIR). Many MIR tasks can be solved by training a classifier over a set of features. For MIR tasks based on music audio, it is possible to extract features from the audio with signal processing techniques. However, some musical aspects are hard to extract with simple heuristics. To obtain richer features, we can use ML to learn a representation from the audio. These learned features can often improve performance for a given MIR task. In order to learn interesting musical representations, it is important to consider the particular aspects of music audio when building learning models. Given the temporal and spectral structure of music audio, deep and multi-scale representations are particularly well suited to represent music. This thesis focuses on learning representations from music audio. Deep and multi-scale models that improve the state-of-the-art for tasks such as instrument recognition, genre recognition and automatic annotation are presented.

Esta dissertação apresenta um estudo comparativo de diferentes métodos computacionais de segmentação estrutural musical, onde o principal objetivo é delimitar fronteiras de seções musicais em um sinal de áudio, e rotulá-las, i.e. agrupar as seções encontradas que correspondem a uma mesma parte musical. São apresentadas novas propostas para segmentação estrutural nãosupervisionada, incluindo métodos para processamento em tempo real, alcançando resultados com taxas de erro inferiores a 12%. O método utilizado compreende um estudo dos descritores sonoros e meios de modelá-los temporalmente, uma exposição das técnicas computacionais de segmentação estrutural e novos métodos de avaliação dos resultados que penalizam tanto a incorreta detecção das fronteiras quanto o número incorreto de rótulos encontrados. O desempenho de cada técnica computacional é calculado utilizando diferentes conjuntos de descritores sonoros e os resultados são apresentados e analisados tanto quantitativa quanto qualitativamente. / A comparative study of different music structural segmentation methods is presented, where the goal is to delimit the borders of musical sections and label them, i.e. group the sections that correspond to the same musical part. Novel proposals for unsupervised segmentation are presented, including methods for real-time segmentation, achieving expressive results, with error ratio less then 12%. Our method consists of a study of sound descriptors, an exposition of the computational techniques for structural segmentation and the description of the evaluation methods utilized, which penalize both incorrect boundary detection and incorrect number of labels. The performance of each technique is calculated using different sound descriptor sets and the results are presented and analysed both from quantitative and qualitative points-of-view.