Music Information Retrieval

The paper introduces G3M, a framework that aims to outline
the musical genome through a memetic analysis of large musical databases. The generated knowledge provides meaningful information about the evolution of musical
structures, styles and compositional techniques over
time and space. Researchers interested in music and sociocultural
evolution can fruitfully use the proposed system
to perform extensive inter-opus analysis of musical works
as well as to understand the evolution occurring within the
musical domain.

On-line keyword searching from documents in Chinese tends to use inverted indexing as the main technique, which has its difficulties. Suffix Array is widely used for processing text in Western languages. However, it fails to get widely used in Chinese processing because of the speciality of Chinese. Suffix Array is a powerful tool. However it costs too much space. That is the major bottleneck of suffix Array. A data
structure called Keyword-driven Suffix Array is proposed in this paper for on-line keyword searching from documents in Chinese, based on observation of on-line search pattern and traits of Chinese. Space efficiency is improved a lot using this data structure. When the document database is large enough, space efficiency is improved by about 5/6 using this data structure without sacrificing its time efficiency.

Musical incipits are short extracts of scores, taken from the beginning. The RISM A/II collection contains about half a million of them. This large collection size makes a ground truth very interesting for the development of music retrieval methods, but at the same time makes it very difficult to establish one. Human experts cannot be expected to sift through half a million melodies to find the best matches for a given query. For 11 queries, we filtered the collection so that about 50 candidates per query were left, which we then presented to 35 human experts for a final ranking. We present our filtering methods, the experiment design, and the resulting ground truth. To obtain ground truths, we ordered the incipits by the median ranks assigned to them by the human experts. For every incipit, we used the Wilcoxon rank sum test to com-pare the list of ranks assigned to it with the lists of ranks assigned to its predecessors. As a result, we know which rank differences are statisticall...

Mixing multitrack music is an expert task where characteristics of the individual elements and their sum are manipulated in terms of balance, timbre and positioning, to resolve technical issues and to meet the creative vision of the artist or engineer. In this paper we conduct a mixing experiment where eight songs are each mixed by eight different engineers. We consider a range of features describing the dynamic, spatial and spectral characteristics of each track, and perform a multidimensional analysis of variance to assess whether the instrument, song and/or engineer is the determining factor that explains the resulting variance, trend, or consistency in mixing methodology. A number of assumed mixing rules from literature are discussed in the light of this data, and implications regarding the automation of various mixing processes are explored. Part of the data used in this work is published in a new online multitrack dataset through which public domain recordings, mixes, and mix ...

Flamenco singing is characterized by pitch instability, micro-tonal ornamentations, large vibrato ranges, and a high degree of melodic variability. These musical features make the automatic identification of flamenco singers a difficult computational task. In this article we present an end-to-end pipeline for flamenco singer identification based on acoustic motif embeddings. In the approach taken, the fundamental frequency obtained directly from the raw audio signal is approximated. This approximation reduces the high variability of the audio signal and allows for small melodic patterns to be discovered using a sequential pattern mining technique, thus creating a dictionary of motifs. Several acoustic features are then used to extract fixed length embeddings of variable length motifs by using convolutional architectures. We test the quality of the embeddings in a flamenco singer identification task, comparing our approach with previous deep learning architectures, and study the effe...

The goal of live song identification is to recognize a song based on a short, noisy cell phone recording of a live performance. We propose a system for known-artist live song identification and provide empirical evidence of its feasibility. The proposed system represents audio as a sequence of hashprints, which are binary fingerprints that are derived from applying a set of spectro-temporal filters to a spectro-gram representation. The spectro-temporal filters can be learned in an unsupervised manner on a small amount of data, and can thus tailor its representation to each artist. Matching is performed using a cross-correlation approach with downsampling and rescoring. We evaluate our approach on the Gracenote live song identification benchmark data set, and compare our results to five other base-line systems. Compared to the previous state-of-the-art, the proposed system improves the mean reciprocal rank from .68 to .79, while simultaneously reducing the average runtime per query from 10 seconds down to 0.9 seconds.

The automated extraction of chord labels from audio recordings constitutes a major task in music information retrieval. To evaluate computer-based chord labeling procedures , one requires ground truth annotations for the underlying audio material. However, the manual generation of such annotations on the basis of audio recordings is tedious and time-consuming. On the other hand, trained musicians can easily derive chord labels from symbolic score data. In this paper, we bridge this gap by describing a procedure that allows for transferring annotations and chord labels from the score domain to the audio domain and vice versa. Using music synchronization techniques, the general idea is to locally warp the annotations of all given data streams onto a common time axis, which then allows for a cross-domain evaluation of the various types of chord labels. As a further contribution of this paper, we extend this principle by introducing a multi-perspective evaluation framework for simultaneously comparing chord recognition results over multiple performances of the same piece of music. The revealed inconsistencies in the results do not only indicate limitations of the employed chord labeling strategies but also deepen the understanding of the underlying music material.

The computer-based harmonic analysis of music recordings with the goal to automatically extract chord labels directly from the given audio data constitutes a major task in music information retrieval. In most automated chord recognition procedures, the given music recording is first converted into a sequence of chroma-based audio features and then pattern matching techniques are applied to map the chroma features to chord labels. In this paper, we analyze the role of the feature extraction step within the recognition pipeline of various chord recognition procedures based on template matching strategies and hidden Markov models. In particular, we report on numerous experiments which show how the various procedures depend on the type of the underlying chroma feature as well as on parameters that control temporal and spectral aspects.

The concept of similarity matrices (SMs) has been widely used for a multitude of music analysis and retrieval tasks including audio structure analysis or version identification. For such tasks, the improvement of structural properties of the similarity matrix at an early state of the processing pipeline has turned out to be of crucial importance. In this paper, we present the SM toolbox, which contains MATLAB implementations for computing and enhancing similarity matrices in various ways. Furthermore, our toolbox includes a number of additional tools for parsing, navigation, and visualization synchronized with audio playback. Finally, we provide the code for a recently proposed audio thumbnailing procedure that demonstrates the applicability and importance of enhancement concepts. Providing MATLAB implementations on a website under a GNU-GPL license and including many illustrative examples, our aim is to foster research and education in music information retrieval.

Humans tend to organize perceived information into hierarchies and structures, a principle that also applies to music. Even musically untrained listeners unconsciously analyze and segment music with regard to various musical aspects, for example, identifying recurrent themes or detecting temporal boundaries between contrasting musical parts. This paper gives an overview of state-of-the-art methods for computational music structure analysis, where the general goal is to divide an audio recording into temporal segments corresponding to musical parts and to group these segments into musically meaningful categories. There are many different criteria for segmenting and structuring music audio. In particular, one can identify three conceptually different approaches, which we refer to as repetition-based, novelty-based, and homogeneity-based approaches. Furthermore, one has to account for different musical dimensions such as melody, harmony, rhythm, and timbre. In our state-of-the-art report, we address these different issues in the context of music structure analysis, while discussing and categorizing the most relevant and recent articles in this field.

As a result of massive digitization efforts and the world wide web, there is an exploding amount of available digital data describing and representing music at various semantic levels and in diverse formats. For example, in the case of the Beatles songs, there are numerous recordings including an increasing number of cover songs and arrangements as well as MIDI data and other symbolic music representations. The general goal of music synchronization is to align the multiple information sources related to a given piece of music. This becomes a difficult problem when the various representations reveal significant differences in structure and polyphony, while exhibiting various types of artifacts. In this paper, we address the issue of how music synchronization techniques are useful for automatically revealing critical passages with significant difference between the two versions to be aligned. Using the corpus of the Beatles songs as test bed, we analyze the kind of differences occurring in audio and MIDI versions available for the songs.

Musical information retrieval (MIR) applications have become an interesting topic both for researchers and commercial applications. The majority of the current knowledge on MIR is based on Western music. However, traditional genres, such as Classical Turkish Music (CTM), have great structural differences compared with Western music. Then, the validity of the current knowledge on this subject must be checked on such genres. Through this work, a MIR application that simulates the human music processing system based on CTM is proposed. To achieve this goal, first mel-frequency cepstral coefficients (MFCCs) and delta-MFCCs, which are the most frequent features used in audio applications, were used as features. In the last few years deep belief networks (DBNs) have become promising classifiers for sound classification problems. To confirm this statement, the classification accuracies of four probability theory-based neural networks, namely radial basis function networks, generalized regr...

Subject of this master thesis is the automatic following of a music score through analysis of sound caused by playing a keyboard instrument. This thesis suggests an algorithm embedded in an e-learning application developed to learn to play a keyboard instrument. Well-known approaches based on Onset Detection and Pitch Chroma features (a.k.a. Pitch Class Profiles) are used to detect played tones and chords in order to trace their chronological sequence and follow them on a virtual musical score. The e-learning application including the Score Following algorithm are based in the web domain and therefore can be executed in modern web browsers.

Özetçe— Bu çalışmada 6 Klasik Türk Müziği makamının derin anlama ağ yapısı ile sınıflandırılması üzerine çalışılmıştır. Öznitelik olarak Mel ve delta Mel frekans kepstral katsayıları, sınıflayıcı olarak ta Derin Anlama Ağları kullanılmıştır. Elde edilen sonuçlar daha önceki çalışmalarda bulunan sonuçlar ile karşılaştırılmıştır. Çalışmada en yüksek başarım Mel frekans kepstral katsayıları ve derin anlama ağı ile %92.70 olarak elde edilmiştir. Yapılan çalışmanın sonucunda literatürde mevcut olan çalışmalardan daha başarılı sonuçlar elde edilmiştir. Anahtar Kelimeler — Klasik Türk Müziği; Mel Frekans Kestral Katsayıları; Derin Anlama Ağları. Abstract—Through this work, six most known Classic Turkish Music Makams (CTM) were classified by using Deep Belief Networks (DBN). The Mel and delta-Mel Frequency Cepstral Coefficients (MFCC) were used as features. The best correct recognition ratio was obtained as 92.70% by using Deep Belief Networks and Mel frequency epstral coefficients. This re...

We are interested in the Music Information Retrieval task that aims at predicting whether a given song will be a commercial success prior to its distribution, based on its audio. This is the topic of what is commonly referred to as ‘Hit Song Science’ (HSS). The underlying assumption behind HSS is that popular songs are similar with respect to a set of features that make them appealing to a majority of people. These features could then be exploited by learning machines in order to predict whether a song will rise to a high position in the chart.

"Carter's music poses struggles of opposition, for instance in timbre (Double Concerto), space (String Quartet No. 3) or pulse (String Quartet No. 5). His preference for the all-interval tetrachords, 4–Z15 [0, 1, 4, 6] and 4–Z29 [0, 1, 3, 7], is also well known. From these facets of Carter's music, I develop a narrative interpretation of his Petrarch sonnet–inspired solo flute piece, Scrivo in Vento (1991). Specifically, I forge narrative pathways by imagining the two tetrachords as active agents opposed in competition. Previous Scrivo analyses (Capuzzo 2002; Childs 2006) stress continuity by revealing Q-transforms and common-note voice leading between the tetrachords. While acknowledging such features, my analysis emphasises oppositional struggle by tracing the tetrachords as separate entities which cooperate and conflict as they manoeuvre to outdo each other.

The analysis advances three theses: (1) it guides listening to and reading Scrivo in a way which resonates with Carter's concern for the aesthetics of oppositional struggle, his choice of a sonnet as inspiration and his affinity for all-interval tetrachords; (2) it shows that music-analytical detail can be organised into dramatic narratives by (a) projecting dramatic roles onto categories asserted by a formal theory and (b) treating the formal theory's relations metaphorically as actions performed by each role as the musical work unfolds; and (3) it shows how detailed pc-set analysis can support a Heraclitean view of music: a flux of opposing forces seeking and resisting unity."

Most of existing audio fingerprinting systems have limitations to be used for high-specific audio retrieval at scale. In this work, we generate a low-dimensional representation from a short unit segment of audio, and couple this fingerprint with a fast maximum inner-product search. To this end, we present a contrastive learning framework that derives from the segment-level search objective. Each update in training uses a batch consisting of a set of pseudo labels, randomly selected original samples, and their augmented replicas. These replicas can simulate the degrading effects on original audio signals by applying small time offsets and various types of distortions, such as background noise and room/microphone impulse responses. In the segment-level search task, where the conventional audio fingerprinting systems used to fail, our system using 10x smaller storage has shown promising results. Our code and dataset will be available.

Grouping songs together, according to music preferences, mood or other characteristics, is an activity which reflects personal listening behaviours and tastes. In the last two decades, due to the increasing size of music catalogue accessible and to improvements of recommendation algorithms , people have been exposed to new ways for creating playlists. In this work, through the statistical analysis of more than 400K playlists from four datasets, created in different temporal and technological contexts, we aim to understand if it is possible to extract information about the evolution of humans strategies for playlist creation. We focus our analysis on two driving concepts of the Music Information Retrieval literature: popularity and diversity.

The International Journal of Software Engineering & Applications (IJSEA) is a bi-monthly open access peer-reviewed journal that publishes articles which contribute new results in all areas of the Software Engineering & Applications. The goal of this journal is to bring together researchers and practitioners from academia and industry to focus on understanding Modern software engineering concepts & establishing new collaborations in these areas.

Bu çalışmada Türk makam müziğinin ezgisel boyutuna yönelik interaktif bir eğitim programı oluşturulması hedeflenmiştir. Geliştirilen araç kapsamlı bir eğitim programı değildir; belirli bir makamı öğrenmek için kullanıcıya referans kayıtlar ve kayıtlarla ilgili öğrenime yardımcı bilgiler ve araçlar sunan bir programdır.

Program, bünyesinde ses kayıtlarının baskın ezgilerini, entonasyonlarını ve icra edilen karar perdelerinin frekanslarını otomatik analiz eden araçlar barındırmaktadır. Aynı zamanda seçilen ses kayıtları ilgili notalarla otomatik eşlenmiştir. Dolayısıyla kullanıcı, referans kayıtlardaki icraların bir bölümünü seçebilir ve çalıştığı bölümün notasını takip edebilir. Tez kapsamında kullanıcıya interaktif bir çalışma ortamının sağlanabilmesi hedeflendiğinden, ses kayıtlarıyla ilgili notaları görselleştiren ve seslendiren araçlar da çalışma kapsamında geliştirilmiştir. Öğrenime yardımcı tüm bilgiler, bir çizimsel arabirim ile kullanıcıya sunulmaktadır. Tez kapsamında yapılan çalışmalar doğrultusunda baskın ezgi analizi ve otomatik karar perdesi frekansı tespiti konuları için yeni yöntemler geliştirilmiştir. Geliştirilen yöntemlerin verimliliği oluşturulan deney veri kümeleri yardımıyla değerlendirilmiştir. Bu çalışma çerçevesinde geliştirilen tüm araçlar ve derlenen veriler, akademik araştırmalar için kullanıma açılmıştır.

Record companies invest billions of dollars in new talent around the globe each year. Gaining insight into what actually makes a hit song would provide tremendous benefits for the music industry. In this research we tackle this question by focussing on the dance hit song classification problem. A database of dance hit songs from 1985 until 2013 is built, including basic musical features, as well as more advanced features that capture a temporal aspect. A number of different classifiers are used to build and test dance hit prediction models. The resulting best model has a good performance when predicting whether a song is a ‘top 10’ dance hit versus a lower listed position.

Time-aligned lyrics can enrich the music listening experience by enabling karaoke, text-based song retrieval and intra-song navigation, and other applications. Compared to text-to-speech alignment, lyrics alignment remains highly challenging, despite many attempts to combine numerous sub-modules including vocal separation and detection in an effort to break down the problem. Furthermore, training required fine-grained annotations to be available in some form. Here, we present a novel system based on a modified Wave-U-Net architecture, which predicts character probabilities directly from raw audio using learnt multi-scale representations of the various signal components. There are no sub-modules whose interdependencies need to be optimized. Our training procedure is designed to work with weak, line-level annotations available in the real world. With a mean alignment error of 0.35s on a standard dataset our system outperforms the state-of-the-art by an order of magnitude.

Most of existing audio fingerprinting systems have limitations to be used for high-specific audio retrieval at scale. In this work, we generate a low-dimensional representation from a short unit segment of audio, and couple this fingerprint with a fast maximum inner-product search. To this end, we present a contrastive learning framework that derives from the segment-level search objective. Each update in training uses a batch consisting of a set of pseudo labels, randomly selected original samples, and their augmented replicas. These replicas can simulate the degrading effects on original audio signals by applying small time offsets and various types of distortions, such as background noise and room/microphone impulse responses. In the segment-level search task, where the conventional audio fingerprinting systems used to fail, our system using 10x smaller storage has shown promising results. Our code and dataset will be available.

Access to digital music collections is nowadays facilitated by content-based methods that allow the retrieval of music on the basis of intrinsic properties of audio, in addition to advanced metadata processing. However, access to ethnic music remains problematic, as this music does not ...

This paper presents TarsosDSP, a framework for real-time audio analysis and processing. Most libraries and frameworks offer either audio analysis and feature extraction or audio synthesis and processing. TarsosDSP is one of a only a few frameworks that offers both analysis, processing and feature extraction in real-time, a unique feature in the Java ecosystem. The framework contains practical audio processing algorithms, it can be extended easily, and has no external dependencies. Each algorithm is implemented as simple as possible thanks to a straightforward processing pipeline. TarsosDSP's features include a resampling algorithm, onset detectors, a number of pitch estimation algorithms, a time stretch algorithm, a pitch shifting algorithm, and an algorithm to calculate the Constant-Q. The framework also allows simple audio synthesis, some audio effects, and several filters. The Open Source framework is a valuable contribution to the MIR-Community and ideal fit for interactive MIR-applications on Android.

Au cours des dernières années, l'apprentissage profond s'est établi comme la nouvelle méthode de référence pour les problèmes de classification audio et notamment la reconnaissance d'instruments. Cependant, ces modèles ne traitent généralement pas la classification de modes de jeux avancés, question pourtant centrale dans la composition contemporaine. Les quelques études réalisées se cantonnent à une évaluation sur une seule banque de sons, dont rien n'assure la généralisation sur des données réelles. Dans cet article, nous étendons les méthodes de l'état de l'art à la classification de modes de jeu instrumentaux en temps réel à partir d'enregistrements de solistes. Nous montrons qu'une combinaison de réseaux convolutionnels (CNN) et récurrents (RNN) permet d'obtenir d'excellents résultats sur un corpus homogène provenant de 5 banques de sons. Toutefois, leur performance s'affaiblit sensiblement sur un corpus hétérogène, ce qui pourrait indiquer une faible capacité à généraliser à des données réelles. Nous proposons des pistes pour résoudre ce problème. Enfin, nous détaillons plusieurs utilisations possibles de nos modèles dans le cadre de systèmes interactifs.

تقدم هذه الورقة نظاما حاسوبيا قادرا على تحليل واستخلاص خصائص الموسيقى العربية لالة الناي العربي بهدف أتمتة التدوين الموسيقي لها. وقد تم استخدم الناي كنموذج لاجراء هذا البحث بهدف تمكين المهتمين في مجال الموسيقى من تحليل واستخلاص النغمات الموسيقية لآلة الناي وتدوينها حاسوبيا. وجدير بالإشارة هنا بان أتمتة التدوين الموسيقي لآلة الناي هو الأكثر صعوبة مقارنة بالآلات الغربية والآلات العربية الأخرى. ويعود السبب في ذلك إلى اختلاف الموسيقى العربية عن نظيرتها الغربية بالأبعاد الموسيقية (Musical intervals)، حيث يعد نصف البعد (Semitone) هو أقل مسافة لحنية ممكنة بين نغمتين في الموسيقى الغربية مقارنة بربع البعد (Quarter tone) في الموسيقى العربية. وكذلك تعتبر عملية التدوين الموسيقي لآلة الناي أكثر صعوبة مقارنة بالآلات العربية الأخرى لكونها بسيطة التركيب وبالتالي فإنها أكثر عرضة لانحراف الضبط المصنعي (Tune deviation).  وتختلف آلة الناي  عن آلات التخت العربي الأخرى بعدم القدرة على تحديد بداية النغمة اعتمادا على تغير شدة الصوت. تم تصميم وبناء نموذج لنظام التدوين الحاسوبي لآلة الناي وتم اختبار كفاءته عمليا من خلال استخدام مخرج النظام وهو ملف ميدي (MIDI File) كمدخل لنظام سيبليوس(Sibelius). وقد بينت نتائج الاختبارات العملية بان النظام المقترح يعمل بدقة تزيد عن 90%.

Esta tese investiga as características e condições de produção e comunicação da playlist musical automática em plataformas de streaming. Ao definir a playlist como um formato, destaca a importância de se estudar o emaranhado de valores culturais, políticos e sociais nas decisões que prescrevem práticas de consumo musical e dão forma aos artefatos com os quais se tem acesso à produção cultural contemporânea. Frente a um objeto de estudo refratário a pesquisas independentes, emprega uma metodologia experimental que mistura investigação de documentos empresariais, editoriais e arquivos eletrônicos com revisões bibliográficas sistematizadas de produções acadêmicas sobre o conceito de playlist e sobre a recomendação automática musical. Complementa-se a análise documental com entrevistas e análise das ofertas musicais dentro dos aplicativos dos serviços. A pesquisa, de cunho qualitativo, busca a emergência de temas e conceitos chave para compreender as maneiras como se constróem as playlists e como são divulgadas e avaliadas por aqueles que as produzem, programadores de sistemas de recomendação e editores musicais. O trabalho assim, defende que a playlist automática é construída com base em um ouvinte imaginado que busca uma experiência individual e personalizada e uma filosofia de programação de engenharia das experiências de audição musical. Além disso, o formato da playlist pressupõe a impossibilidade de separação entre expressões automáticas e editoriais, uma vez que as duas se misturam em sua apresentação pelas plataformas de streaming e nos cotidianos de produção, orientados pela plataformização dessas indústrias de difusão e distribuição musical digitais. Levanta-se, por fim, as possíveis consequências dessa forma de abordar a prescrição musical para a promoção da diversidade musical e para a própria experiência musical na sociedade contemporânea.

A musical performance renders an acoustic realization of a musical score or other representation of a composition. Different performances of the same composition may vary in terms of performance parameters such as timing or dynamics, and these variations may have a major impact on how a listener perceives the music. The analysis of music performance has traditionally been a peripheral topic for the MIR research community, where often a single audio recording is used as representative of a musical work. This paper surveys the field of Music Performance Analysis (MPA) from several perspectives including the measurement of performance parameters, the relation of those parameters to the actions and intentions of a performer or perceptual effects on a listener, and finally the assessment of musical performance. This paper also discusses MPA as it relates to MIR, pointing out opportunities for collaboration and future research in both areas.

A 24-dimensional model for the ‘harmonic content’ of pieces of music has proved to be remarkably robust in the retrieval of polyphonic queries from a database of polyphonic music in the presence of quite significant noise and errors in either query or database document. We have further found that higher-order (1st- to 3rd-order) models tend to work better for music retrieval than 0th-order ones owing to the richer context they capture. However, there is a serious performance cost due to the large size of such models and the present paper reports on some attempts to reduce dimensionality while retaining the general robustness of the method. We find that some simple reduced-dimensionality models, if their parameter settings are carefully chosen, do indeed perform almost as well as the full 24-dimensional versions. Furthermore, in terms of recall in the top 1000 documents retrieved, we find that a 6-dimensional 2nd-order model gives even better performance than the full model. This represents a potential 64-times reduction in model size and search-time, making it a suitable candidate for filtering a large database as the first stage of a two-stage retrieval system.

There is a small but growing body of research addressing compositional process. The majority of this research is qualitative in nature and requires intrusive data collection methods. What quantitative research does exist is limited in scope due to a lack of effective data collection tools. Whether qualitative or quantitative, the analysis of compositional process data with current methods is labor-intensive and scales poorly.
To address these concerns a custom piece of notation software, Thema, was created. Thema automatically tracks composers’ actions within the software environment and stores the date in a MySQL database; the database is integrated into the program and plays an active role in its operations. Entities in the score, such as notes, measures, and dynamics, are tracked across time using permanent identifiers. Thema also records the state of the notation environment as the composer works including viewable area, selections in the score, playback, and continuous MIDI keyboard input. Additionally, Thema records a screen capture for every action performed in the program.
This document describes the architecture and implementation of Thema as well as its usage in a study of ten graduate-level composers. To examine the study’s data, a custom suite of visualizations, Omaggio, was created. Using these visualizations, the development of musical structure over time is investigated in detail; a case study of local problem-solving behavior in the context of improvisational and score data is also presented.
This work has significant implications for compositional process research, composition pedagogy and musicology, and illustrates the power of combining automatic data collection with a rich data model. This new method of data collection is non-intrusive, anonymous, and practical for regular use by composers. The analytic tools presented herein integrate keyboard and score data and could be helpful to composers during the process of composition. Additionally, these tools could serve a role in future qualitative studies by reducing organizational overhead and providing search capabilities. To facilitate further research, the musical works from the study are released under a Creative Commons license and the compositional process data set will be accessible to researchers.

We present AutoLeadGuitar, a system for automatically generating guitar solo tablatures from an input chord and key sequence. Our system generates solos in distinct musical phrases, and is trained using existing digital tablatures sourced from the web. When generating solos AutoLeadGuitar assigns phrase boundaries, rhythms and fretboard positions within a probabilistic framework, guided towards chord tones by two user­ specified parameters (chord tone preference during and at the end of phrases). Furthermore, guitar-specific ornaments such as hammer-ons, pull-offs, slides and string bends are built directly into our model. Listening tests with our model output confirm that the inclusion of chord tone preferences, phrasing, and guitar ornaments corresponds to an increase in user satisfaction.