My first answer is "not much". The first lesson of the statistical approach is the confirmation of music-theoretical common sense: there are vast stretches of music where chords and nonharmonic tones
behave exactly as the textbooks say they should. This poses a challenge to a whole host of radical and revisionist theories of musical structure. I consider three cases: von Hippel and Huron's denial of "post- skip reversals". Schoenberg's skepticism about "nonharmonic tones" (recently echoed by Ian Quinn), and Schenker's rejection of local harmonic constraints in functional tonality.
My second answer is "everything". For corpus studies can reveal numerous instances that fundamentally controvert standard music-theoretical assumptions: that Renaissance nonharmonic tones decorate a fundamentally triadic or consonant substrate, that functional-harmonic behavior is
systematic. My talk will focus on how to reconcile these two answers.

Although corpus studies have become quite common in musicology and music theory, an implicit divide still exists between the questions addressed with distant readings and the questions thought to require close reading. This schism is made all the more complicated by instances in which the musical "data" being examined is missing, incomplete, or apocryphal. We are forced to examine authorship without primary sources, performance histories based on only a handful of early recordings, and aspects of orality, memory, and transmission with nothing but written evidence. Here I demonstrate how such distant readings built upon instances of ambiguity are nevertheless informative of close readings. While discussing the explanatory limits of corpus methods, I also argue that these limits should not be seen as disqualifiers and that the ambiguities inherent in such questions often benefit from such empirical and corpus-based method.

Musical form is an essentially top-down concept, concerned with the organization of parts within a whole, and their interrelations with one another. Yet in many musical forms, most particularly in sonata form, theory trails practice by many decades, suggesting that form emerged not through top-down theoretical dictates, but rather through a bottom-up process of self-organization. Hence, top-down rationalizations of form, although valuable from a synchronic perspective, only rarely do justice from a diachronic one. That is, they may explain what a form was at a given point, but not how it came to be that way. Here I demonstrate how corpus studies can support such diachronic sensitivity, not only by providing a chronologically sensitive notion of common-practice, but also by illuminating the dynamic nature of form as a system: tension between elements, processes of change and instability, and islands of stasis within a system that is overall in constant flux. Through the dynamics of the system of binary, and later, sonata form, I argue that these tensions explain the use of specific formal “stereotypes” at particular times in the 18th century. While I advocate the advantages of such an approach, I will also discuss the limitations of the corpus-based technique, including what cannot be measured, and what cannot be proven.

What are the primitives from which musical structure is built? According to reductive theories such as Schenkerian analysis, one fundamental principle is the recursive elaboration and complexification of simpler structures using ornamentations such as neighbor and passing notes. In this talk, an attempt is presented to put this intuition on a solid formal foundation using a graph-based representation of musical structure. The resulting system is able to express hierarchical relationships between notes from the background to the surface of a piece. In addition to the basic model, extensions that capture simultaneity and harmonic properties will be discussed.

Easily creating, using, and sharing music annotation data are challenges for digital
musicology. How can we formalize music annotation elements, encode them, and use
them in systematic musicology approaches? I will present Dezrann (www.dezrann.net), a
web platform to share music analysis for corpus-based musicology and music information
retrieval (MIR) research developed by Emmanuel Leguy in the Algomus team. Music is
presented either on continuous staves or waveforms. The music analyst creates, edits and
moves labels on or around the score, then saves and shares her analysis. Design goals
were to be open-source, multi-platform (no installation, web platform, both desktop and
mobile) and simple to use. I will detail how we used the platform while annotating sonata
forms in Mozart’s string quartets and present ongoing work on real-time collaborative
annotation.

Tonality is a central cognitive category of Western Classical music. Previous empirical research presupposes that musical pieces are either in the major or the minor mode. That assumption, however, does not entirely account for the cultural complexity and historical development of tonality. In this talk, I present a novel model that does not make prior assumptions about the number and structure of modes but rather determines them via unsupervised Bayesian inference. I propose a geometric and a probabilistic model of tonality and its essential components key and mode. Using a dataset of approximately 13,000 pieces, I show the emergence of the major and the minor mode in the Baroque period, Atheir stabilization in the Classical era, and their subsequent divergence during the course of the 19th century. Finally, I give an outlook on how the inferred modes can be arranged to a mode hierarchy.

Pitch-class distributions are widely used in corpus-based Computational Musicology, e.g. to determine the mode or key of musical pieces, or to infer style-specific characteristics (De Clerq & Temperley, 2011; Broze & Shanahan, 2013). Unfortunately, research is often confined by specific representations of musical pieces due to MIDI-encoded data that renders, for instance, the distinction between enharmonically equivalent pitch classes (e.g. F# and Gb) and between diatonic and chromatic intervals (e.g. m2 and A1) impossible. Moreover, historical changes of these distributions are only recently being studied on a larger scale (Tymoczko, 2011; Albrecht & Huron, 2014; White, 2014; Yust, 2019), while many approaches assume more or less immutable abstract “tone profiles” (Krumhansl, 1990). These shortcomings are addressed by analyzing distributions of “tonal pitch classes” (TPCs; Temperley, 2000) in a large dataset of more than 2000 MusicXML-encoded pieces by 75 composers covering a historical range of almost 600 years.

It is shown that the order on the line of fifths can be inferred both from TPC distributions of pieces as well as diachronically. Moreover, a mathematical model is used to show that composers progressively use third-based tonal relations in an increasingly explorative manner. Thus, the historically changing distributions of TPCs allow exposing large-scale changes in compositional practices. The present approach argues that adequate models for musical notes and pieces, as well as the acknowledgment of historical changes, are indispensable for the computational study of tonality.

This study proposes a novel approach to modeling musical form in tonal music by analyzing pitch-class distributions in symbolic data using a variety of machine learning algorithms. Results suggest that sliding-window histograms, which take the temporal component into account, improve algorithms' performance in assigning form labels to pieces. Pitch-class histograms were extracted from major-mode solo piano pieces by W. A. Mozart and L. v. Beethoven according to two methods: whole-piece histograms and sliding-window histograms. In the latter method, richer features were obtained by calculating histograms for 90 partially overlapping windows per piece. Unsupervised clustering employing Gaussian Mixture Model (GMM) yields good results in separating sonata-form movements from non-sonata-form movements. Supervised learning applying Support Vector Machine (SVM) manages to separate data-points according to three distinct form labels: sonata form, minuet/scherzo form, and variation form. In both analyses, using sliding-window histograms significantly improves results. Finally, we discuss the adequacy of sliding-window histograms for modeling further features such as a composer's personal style.

A method is presented for detecting classical cadences using a finite state machine.

The state-machine detector (also called finite-state automaton) can be seen as a combination of a decision tree and a finite Markov chain. The algorithm does not require training data and uses mostly harmonic and melodic progressions to move between what I define as Cadence- Related States.

The state machine is based on the conception that, during the listening experience we shift between states as a reaction to the musical stimuli, which is what ultimately generates our comprehension as well as our expectation of next events. Conceptually, these states are related to musical perception as modeled by Meyer and Narmour and are also in accordance with the generative model of Lerdahl and Jackendoff.

The suggested finite states are: Idle, Cadence Expected, Cadence Inevitable, Cadence

Avoided, Perfect Authentic Arrival, Imperfect Authentic Arrival. The triggers to transfer between cadence-related states are events that originate from the previous state and some current input (such as notes or chords). The input events vary but are based on the harmony, the bass, and the soprano.

The algorithm is implemented in python code using the Music21 library. The state machine is evaluated on a database of 44 Haydn string quartets used in prior research and shows promising results of over 80% recall rate. The state machine approach enables an investigation and discussion on the performance of the detector with regards to the known definitions of cadence and cases of exceptions are observed.

Transition tables (TTs) are a popular and useful tool in music research, especially for dealing
with chords. It is not clear, though, how this representation interacts with our perception of
chords and the relation between them. On the other hand, 'musical spaces' – another
popular representation – are assumed to be related to the way we process chords, but their
relation to the corpus-based analysis is usually unclear.
I will show that TTs indeed correlate with supposed harmonic spaces and discuss the
corpuses, transformations, and metrics that maximize these correlations. One preliminary
result shows that the correlations differ significantly with respect to the distribution of
consequence chords compared with that of antecedent chords. In the talk, I will furthermore
display associations between correlations of different genres and periods and will suggest
ways to analyze these correlations and the operations and parameters that maximize them.
Moreover, I will use a multidimensional scaling to reconstruct optimal spaces from given
TTs. I will compare the outputs of this procedure for TTs of different styles and eras and will
discuss both similarities and dissimilarities between the reconstructed spaces and previously
suggested spaces.

I would like to briefly present the dataset I have created for my Ph.D. research project, which contains manual analyses of 1001 Israeli pop songs, from the five most recent decades. The songs were chosen from the yearly charts, with equal representation for the two main styles in contemporary Israeli pop: Israeli Pop-Rock and Múziqa Mizraḥít (in Hebrew: "oriental music"). The analyses describe musical features such as form, rhythm, harmony, and timbre, as well as metadata and lyrics. In addition, the dataset includes the audio itself, and audio features extracted with the Librosa package for Python.

The main research questions have to do with style differentiation over time, style comparison and complexity. While building the dataset, some observable style differences have emerged. Those findings – presented through statistical analysis – were used as a departure point for machine learning methods of data exploration, including PLS and KNN. The ML methods have yielded visualizations and transformations of the data which allowed for a more holistic, "zoomed-out" view of the corpus, exposing long-term chronological stylistic trends which were implicit in the statistical analysis.

The corpus-based exploration of musical structures has gained increasing importance over the past two decades. Apart from studies of tonal harmony in general, a growing number of theoretical, computational, and cognitive studies examine the use of cadences in the “classical” repertoire (e.g., Duane & Jakubowski, 2018; Sears et al., 2018). However, computational research on musical structures is still hindered due to (1) the sparsity of available datasets and (2) the lack of shared standards for encoding music analyses.

In an attempt to address these lacunae, this contribution examines cadences and their harmonic characteristics in a newly created dataset consisting of analyses of all keyboard sonatas by Mozart. This dataset contains expert annotations at two analytic levels: (a) Harmonic analyses providing information on key, chordal root, chord form, chord position, and local voice-leading; and (b) annotations of more than 1,000 cadence tokens based on a cadence typology developed in various music theoretical studies (Neuwirth & Bergé, 2015).

Since algorithmic cadence finding is still a challenging and by no means flawless task, we propose a semi-automated approach, one that combines a labeled dataset (cadence type and endpoint) with implemented heuristics that help delineate cadence instances both from one another and from non-cadential harmonic progressions. This allows us to characterize the frequency of cadence types, their harmonic design, and their probabilistic profiles. Thus, this talk aims to provide an empirically testable basis for the use of cadences in the classical style, drawing on Mozart’s keyboard sonatas as a point of departure for future studies.

Although musicians and non-musicians alike recognize authentic and half cadences (Bigand & Parncutt, 1999; Boltz, 1989; Rosner & Narmour, 1992; Sears, Caplin, & McAdams, 2014), it is unclear how they do so, especially without explicit musical knowledge. In this talk, I will review three corpus studies that examine how tonal music affords human learning and perception of cadences. All studies employ the same basic methodology: first, two musical experts identified all authentic and half cadences in a corpus; then, naïve Bayes classifiers and expectation-maximization (EM) algorithms were used to identify cadences and learn cadential categories, respectively. Since both techniques are Bayesian, they reflect plausible models of cognitive processing (Clark, 2013; Knill & Pouget, 2004). The first study examines harmonic cues in a corpus of Classical string quartets, with harmony being modeled using a set of prototypical distributions of scale degrees. These distributions allow cadences to be classified and learned with above-chance accuracy. The second and third studies assess melodic cues—in particular short, prototypical scale-degree sequences in the melody and bass, which were modeled using both n-grams and profile hidden Markov models. The second study applies this method to the Classical string quartets, whereas the third study uses it on a corpus of Galant solfeggi. With the quartets, melodic cues allow above-chance classification but not above-chance EM. With the solfeggi, neither task yields above-chance performance. This surprising result stems from the highly formulaic nature of these pedagogical pieces, according to follow-up analysis of the solfeggi.

The view that musical passages are capable of expressing their location within musical time
through intrinsic qualities has been put forth by William Caplin (2009). Cognitive studies (e.g.,
Vallières 2009) corroborates this view inasmuch as it pertains to determining the formal functions
(beginning, middle and end) of small segments of music at a local level. However, I propose to
show that on a larger scale, a passage’s intrinsically implied temporality may vary from its actual
location along the piece’s timeline. Current concepts of musical form, such as Hepokoski and
Darcy’s “sonata trajectory” (2006), involve determining temporal positions primarily in relation
to major closural events, such as the “middle caesura,” or the “essential expositional closure.”
Heinrich Christoph Koch’s more flexible set of cadential strategies in a sonata exposition makes
it possible to opt for a multiplicity of closural events in a way that is also capable of accounting
for non-linear scenarios, that is, cases where “later” cadence types precede “earlier” ones
(although Koch nowhere explicitly refers to the possibility of a swapped order of cadences). My
corpus-based investigation into closural events in a selection of sonata expositions by Mozart
suggests that such “non-linear” trajectories correlate with the exposition’s size, meaning that a
“twisted” order of cadences may be construed as a prolongational technique in order to sustain
particularly large exposition sections. Collecting patterns of closural events across a corpus of
sonata expositions may be further used for comparative analysis at a “distant reading” level.
Extracting such closural patterns from given exposition sections involves resolving various
methodological difficulties in identifying and classifying different cadence types. However, in
my “close reading” of the first movement expositions of the “Jupiter” Symphony and the Violin
Sonata K. 301, I will try to show that non-linear closural patterns cannot be regarded separately
from a given movement’s thematic layout, nor from its style and genre.

The repertoire of Trecento madrigals is one of the oldest polyphonic settings of vernacular texts.   The approximately 190 madrigals, of which about 170 are two-voice compositions, were gathered mainly in three large manuscripts that were compiled in two local centers – northern Italy (Codex Rossi [Rs]) and Florence (codex Panciatichiano 26 [Fl] and codex Squacalupi [Sq]). The purpose of this corpus-analysis study is to characterize the styling components of the two-voice Trecento madrigals that originated from the two local centers in search of the composers' individual traits alongside the local style and the period style. In a search for recurring patterns, each madrigal was thoroughly examined for mensural, structural, melodic and motivic traits, both on a single line level and terzetto–ritornello one. The results indicate that the flexible frame of the period style enabled the composer interlace in his work personal characteristics, therefore creating an individual style that overshadowed the local elusive style. The composer's voice is revealed mainly by the tenor characteristics and the various types of cantus-tenor relations. The opening declaration of the madrigal and the articulations of lines serve as individual footprints as well.    

The correlation between specific musical features and the expression of emotions has been the cornerstone in musicological studies, from score to empirical analysis. In DIDONE we are digitizing and studying a total of ca. 4000 opera seria arias on libretti by Pietro Metastasio. The corpus, mostly neglected so far, uniquely puts at our disposal numerous musical versions of a reduced number of texts, with a view to examining how composers conceived specific emotional states to be conveyed through music.

In this paper, I will be presenting a preliminary analysis of ca. 100 settings of 3 different arias for the main female character in Didone abbandonata’s libretto. By discussing how composers handled parameters such as scoring, key, meter, tempo, duration, rhythm, vocal range, and intervallic configuration, I will explore:

- How the manipulation of specific musical parameters correlates with the expression of particular emotions in the corpus;

- To what extent a diachronic-aesthetic explanation of the findings is not wholly satisfactory;

- In which manner big data analysis can be of use to musicology.

This paper ultimately aims to contribute to the Workshop discussion about methodologies in corpus analysis, as well as to offer preliminary results regarding historical strategies for the expression of emotions by compositional means.

Haydn’s forms are often held up as anomalies, as special exceptions to a rule. Upon close examination, sometimes these pieces are truly quirky while other times they display syntax common to Haydn, but uncommon to Mozart and Beethoven (upon whom most theories of Classical Form are based).

This work draws on the methodology of corpus studies to investigate questions such as “How can we describe Haydn’s syntax,” “What compositional approaches differentiate Haydn from Mozart and Beethoven,” and “Aren’t there more elegant ways of describing Haydn’s forms than what we currently have?” The corpora are small, medium, and large formal sections from the last movements of Haydn’s symphonies, and currently, encompass the complete last movements of Symphonies Nos. 65-104. One premise that underlies the collected data is that returns to his opening theme are an essential aspect of Haydn’s style. These restatements reveal the inadequacy of the term monothematicism and are more aptly described by a concept I call thematic saturation.

In this presentation, I first situate my work within research on Haydn’s symphonic last movements (Cole 1982, Bonds 1988, Fisher 1992, Haimo 1995, Galand 1995, Fillion 2012, Neuwirth 2013, Hunt 2014, Inman 2016, among others). Then, after laying out first steps towards a theory of thematic saturation, I describe my methodology, how it has evolved, and what I have learned to this point.

Theories of eighteenth-century concerto form are mostly synchronic, distinguishing between early practices associated with Vivaldi, and late ones, associated with Mozart and the type 5 sonata. However, contemporary descriptions by writers suggest a gradual evolution from one formal practice to the other, driven by two main processes: the growing importance of the solo, and the organization of the movement in large-scale formal units. Thus, this study offers a fresh diachronic corpus-based approach to investigate the transformation of the eighteenth-century concerto in light of these two processes.

This research focused on core ritornello-form elements, common to early and late eighteenth-century practices: tutti and solo alternating textures, harmonic trajectory, and repetition of thematic material from the beginning and end of the first ritornello. I used quantitative measures to determine how these elements changed over time in 210 concerto movements by over a hundred different composers.

The findings indicate a gradual evolution of form. Over time, movements substantially grew through the expansion of the opening ritornello, and the first and last solos. The middle and final orchestral passages were significantly reduced in terms of their relative portion of the movement, eventually losing their independence. Furthermore, the proportional role of the solo dramatically grew, as was its role in spelling out the form through thematic content.

The paper is based on the qualitative and quantitative research of the entire corpus of Schubert’s 435 short dances for piano solo. Some intra-corpus norms deviate from the more general norms in the literature, e.g., more pieces change key or mode than pieces that are monotonal in minor; sectional periods are more common than interrupted periods, and imperfect authentic cadences very often have the fifth in the upper voice rather than the third. Many dances include units that function like complete phrases despite their brevity and minimal tonal motion. Over 90% of the dances have entirely square phrase rhythm, but other devices come to the fore. Often it is difficult to determine whether the piece is simple binary or rounded binary since some elements return but others do not (There are multiple combinations of partial return); Reverse hypermeter (hyper downbeats on even measures) is also common, sometimes for part of the dance only, giving rise to various manners of making the change. Also common are dislocations of motives (motives that change their location within a phrase). All data can be traced with exact amounts of examples.

Vivaldi’s contribution to the consolidation of the new type of tonal hierarchy and pitch structure –viewed retrospectively as the tonality of common practice – is considered one of his principal achievements. I explore the entire corpus of Vivaldi’s compositions (24 instant operatic scores, around 480 concertos, 96 sonatas, numerous sacred pieces, dramatic serenatas, and chamber cantatas), irrespective of their generic and structural distinctions, in order to categorize the resources he generated to enable this extraordinary tonal expansion. By these means, Vivaldi was able to attain an artistically convincing equilibrium between striving toward new goals and consolidating the footholds of tonal stability. Viewed against the background of late Seicento tonality, with its overwhelming compulsion to launch toward a new goal immediately after cadencing, Vivaldi’s style of tonal expansion represents a truly revolutionary achievement. For these purposes, he favors assertive repetitive bass patterns, including ground bass, sequences, shifts into the parallel minor, pedal points, and cadences. I aim to demonstrate the syntactic function of these and related patterns in the context of Fortspinnung periods forming his ritornello, da capo and Kirchenarie, and dance binary movements.

This paper uses corpus research as a means of unlocking musical meaning—specifically the embodied meanings that historical listeners familiar with the danced minuet might have attributed to minuet music outside the ballroom. Analysts past and present have discussed unusual metrical features in non-functional minuets, in which listeners may feel “like stepping into a hole” (Mirka 2009) or where “the feet cannot settle into the prescribed steps” (Vogler 1778). My contribution to this type of inquiry uses corpus analysis to complement such individual readings in two ways. First, by examining the metrical properties of functional minuets and gaining a better understanding of what constitute norms and deviations thereof. Second, by analyzing the typicality of compositional procedures that Haydn and Mozart deployed in minuets that were not intended to be danced.

I identify features regarding hypermeter and phrase structure that are common in certain genres and unusual in others and provide a taxonomy of expressive strategies involving dance music relations to interpret such differences. For example, twelve-bar phrases and two-bar addendums appear with significantly higher frequency in symphonic minuets than in those written for the chamber or the dance floor. Although these procedures deny four-bar hypermeter—often considered characteristic of the minuet (Love 2015, McClelland 2006)—they provide an overt imitation of dance steps. I interpret such musical analogies as a response to the aesthetic demands and public nature of the genre, only apparent when examining the corpus of symphonic minuets as a group.

In order to delineate a perceptual model of musical expressiveness in solo violin performance, we found appropriate to use a description based on Expressive Musical Terms (EMTs). These qualifiers, although directly related to the musical language, have rarely been utilized in the field of music cognition. For this purpose, we build an audio corpus of short musical excerpts from the violin classic repertoire, each of them performed by a violinist according to a few EMTs and examine the listeners’ perception of these excerpts. To select these EMTs, we require they arise as centroids of clusters making a partition of a data set of 55 EMTs associated with a broad range of musical characters. To make a partition based on the perception of professional string players, we performed two experiments. In the first one, the participants were asked to organize the data set in a two-dimensional plane in such a way that proximity reflects similarity. Using a minimization procedure, we found that a satisfactory partition is provided by four clusters with centroids associated with tenderness, happiness, anger, and sadness. Through a Multi-Dimensional Scaling analysis, these clusters are located in the four quarters of the valence-arousal plane of the circumplex model of effect developed by Russell (1980). In a second experiment, the participants had to rate directly the data set elements according to valence, arousal, extraversion, and neuroticism, using 7-level Likert scales.  We found strong consistency between the two experiments. In terms of the related positive-negative activation parameters, introduced by Watson and Tellegen (1985), we obtained a significant correlation between positive activation and extraversion and between negative activation and neuroticism. This demonstrates that these relations, previously observed in personality studies by Watson & Clark (1992), extend to the musical field.

Music theorists have often associated tonal functions or progressions with non-auditory domains, including physical space (e.g., tonal center), physical motion (leading tone, cadence) or visual brightness (“dark” chromatics). Yet, do such cross-modal connotations actually affect the perception and cognition of tonality – the ways tonal progressions are “felt” by listeners? Correspondingly, are such connotations consistently expressed when music is matched with texts or images? In this presentation, we report a series of experiments addressing the first question and suggest how corpus studies of music associated with text or image may address the second. We investigated cross-modal correspondences (CMC)  between tonal stability and several non-auditory domains, including visual brightness, physical size, spatial location (high/low, right/left), and visually-depicted emotion. Correspondences were examined in both musicians and non-musicians, using both explicit and implicit measures. On the explicit test, participants heard a tonality-establishing cadence followed by a probe-tone and matched each probe to one of several visual stimuli (e.g., circles varying in brightness). On the implicit test, we applied the Implicit Association Test (IAT) to auditory (tonally stable or unstable sequences) and visual (e.g., bright or dark circles) stimuli. Participants reliably associate tonally stable notes with brighter, higher, and happier visual stimuli. These associations affected both musically-educated and musically-naïve listeners, both explicitly and implicitly, thus suggesting that they depend only partially on conceptual musical knowledge. The results demonstrate how a seemingly abstract musical schema may establish concrete connotations to non-auditory perceptual domains, thus suggesting a hitherto unexplored path associating syntax and connotative meaning in music. We propose follow-up corpus research of these intriguing correspondences, which may systematically examine whether the degree of tonal stability and closure affects music-text and music-image relationships in genres such as lieder, opera, film music, or music videos.

It might seem that there’s nothing left to say about the classical cadence: William Caplin
devoted a foundational, 68-page article to the perfect authentic cadence in 2004; in 2015,
Markus Neuwirth and Pieter Bergé went one better and published an entire book. What more
could there be to do?
I will nonetheless argue: (1) that a redefinition of cadential function is necessary in order to
resolve the tension in Caplin’s theory between a conceptual commitment to “intrinsic” formal
function and de facto appeals to contextual factors; (2) that the correct redefinition is in terms
of three-voice contrapuntal models; (3) that this redefinition leads in turn to a helpful
clarification of the concept of the “half cadence”; and (4) that the resulting perspective is in
principle extensible to medial and initiating functions as well, in a way that promises a fusion of
form-functional thinking with schemata or Satzmodelle.
In methodological terms, I will emphasize the usefulness of quantitative work with carefully
defined and appropriately selected corpora over the reliance on “intuitive” statistics that has
been the norm in music-theoretical research.
All of these points will be illustrated with examples and results drawn from my ongoing work on
form in Mozart’s music, with the corpora in question being (1) the piano sonatas; (2) the string
quintets; (3) a set of 180 works sampled randomly from the Neue Mozart Ausgabe; and (4) the
operatic and concert arias from Idomeneo to La Clemenza di Tito and Die Zauberflöte.

Human language is a complex phenomenon that is far from being well understood. Letters are combined to form words, words combine to phrases and sentences, and these, in turn, combine to form paragraphs and document. The structure of this combination is constrained by many different factors: morphological, syntactic, semantic, discourse, and pragmatics. Language modeling (LM) is the task of statistical modeling of this process, in an attempt to predict the next symbol in a sequence. After presenting the challenges in the language modeling task, I will describe some recent advances---which revolutionized the field---and which are based on non-markovian modeling using recursive and attentive neural networks. While the topic is somewhat technical, the talk won't be.