Platform(s)

The OpenScore Lieder Corpus has explicitly set out to meet the criteria outlined above, providing a resource that scholars, teachers, and musicians can use for any purpose. The corpus now offers over 1,200 high-quality encodings of songs that can be viewed, played, transposed, and more either directly online at https://musescore.com/openscore-lieder-corpus/sets or once downloaded in any of several standard formats. 11 Usage statistics show that users do indeed engage with these scores hundreds of times per day (via views, plays, downloads, and more).

We reached this level of engagement by factoring it in from the outset: this is not exclusively a research dataset, but a score collection for which musicological research is one of several use cases. That simple fact informs every decision, including the primary host platform (musescore.com), which is set up for engagement with a wide musical community. Indeed, it is our attention to FAIR principles for the research side that has been slower to develop: downloading individual scores from musescore.com is easy, but not currently possible at corpus scale. To that effect, we have now developed GitHub mirror(s) of the entire corpus, 12 which in turn form the basis of an analytical corpus and a set of teaching resources to which we will return below.

Ultimately, an ideal platform would support all human and computational searches to find the desired content easily, whether that's a PDF, a symbolic score, a recording (audio or video), or metadata about pieces (including information about movements and sections) and would even support diverse search terms for slippery concepts like genre and analytical properties. Sites like Musicbrainz.org and Doremus.org offer promising approaches to these matters of connected 'music ontology', but we no single platform currently offers the full package. For now, the best we can do is to host scores on complementary platforms (musescore.com and GitHub in our case) 13, with clear and consistently structured metadata (for the mostly unequivocal items like title, composer, lyricist, and catalogue number), and with external links to collections like wikidata.org. This serves both the findability of those scores now and also the prospective incorporation of the corpus into future, more richly annotated and interconnected databases.

Editorial 'Policy' and Recruitment

Any corpus-building effort is subject to the limited resources available and to a tradeoff between quantity and quality. As such, the 'best' approach depends on the project's priorities, making a commitment to FAIR and open principles relevant even here. For the OpenScore Lieder Corpus, we have elected to copy specific, identifiable, public domain editions and we provide a link to that source for the sake of transparency and easy comparison. We aimed for reliable transcription quality, correcting only flagrant errors, and stopping far short of a scholarly edition. If we were optimizing for academic reward metrics such as publication, investing so much time in the scores' quality without also producing a product recognizable as a scholarly edition would be among the worst ways to proceed. We saw the problem differently: in order to maximize the number of scores reliably reusable, and to make best use of our amateur volunteers, it was important not to take on anything truly editorial.

That said, one cannot simply decide to 'not do' editing as each source poses specific problems in terms of a balance between fidelity to the reference source and consistency in the new corpus. In principle we have sought to emulate the original as closely as possible, with exceptions for two types of notational devices that affect openness. Firstly, guaranteeing openness to modern musicians requires the 'updating' of elements that have fallen decisively out of common usage. Examples include now-rare clefs or the positioning of pedal indications anywhere other than below the bottom staff. Secondly, we replaced notations which create undue complication in the encoded version, such as the splitting of a single measure across a system break. A representation of this is possible, but only with hidden, non-standard elements (two measures, of which only one is counted) which make conversions, interoperability and other computational processes more difficult. This comes back to balancing the prospective use cases. The measure-split may be a more direct reflection of the source and perhaps better for use in training optical musical recognition software (hereafter OMR), but is much worse for most other computational purposes and is also increasingly unfamiliar to musicians today.

To make these kinds of judgement calls effectively, we need to understand different users' needs. We can arrange consultations to this effect, but the most direct and meaningful engagement is to be had by involving a range of users in the building of the corpus itself, very much including amateurs alongside professional musicians and musicologists. Moreover, apart from bringing complementary priorities and insights, we have found that some of the most committed and successful contributors have begun as musical novices, perhaps because the tasks are proportionately more interesting and valuable for personal, musical development at that stage. 14 With a little encouragement from more experienced members of the team, we have helped these contributors to develop their own skills (a worthwhile service it itself) and benefitted from the transcriptions they subsequently produce.

Platforms with an 'open to all' policy and social-media-esque stylings can help to build this kind of collective transcription community by gathering like-minded people together and promoting all-important personal connections. This has been central to the Lieder Corpus, both for inviting contributors to work on new scores, and (in some cases) for adopting existing ones by reaching out to the transcriber via the site. Where those existing scores are already released under a suitable license, we can thank the transcriber for their work and get to work on any necessary improvements immediately (either separately, or in collaboration with the original transcriber as they prefer). Where the license is not suitable, often the transcriber has simply not considered this aspect: on several occasions, a short exchange has confirmed their preparedness to contribute their work. In some cases, this has been the initial moment of recruitment for highly productive and valuable contributors. This speaks to the benefits of personal contact and open channels of communication with (potentially) interested parties. It also affects the best-practice in workflow.

Workflow

Coordinating multi-author transcription projects requires careful coordination and distribution of tasks. In the Lieder Corpus we have been able to keep this reasonably simple: we take every song as an atomic unit, allocating exactly one transcriber and one (different) reviewer to each, 15 and keep track of progress on a spreadsheet viewable by all but editable only by the review team. More complex projects than ours may require a more sophisticated version of this process, for instance, distributing the task of transcribing among different parties according to their experience and time availability, ranging from simple, one-click reCAPTCHA-style tasks ('do the pitches in these two score excerpts match?') to expert editorial judgement calls. There are promising ideas for implementing this, 16 which if combined with McGill's forthcoming mass OMR conversion of the entire IMSLP catalogue, could be a strong basis for coordination and potentially a watershed moment in the score encoding effort.

That said, we would do well to synthesize project management efficiency with what contributors will enjoy doing. While there is a great enthusiasm for encoding, there seems to be less of an appetite for correcting existing encodings. This may be partly for want of a coordinated system and lack of clarity over licenses, but it would seem to run deeper. For instance, we should not expect wholesale enthusiasm for a set-up in which contributors are expected to fix OMR. 17 OMR has advanced to a point that it is no longer so easy to dismiss revising it as being more time-consuming that transcription from scratch, and yet that argument persists. To engage with that viewpoint, we need to establish whether it has simply become a convenient distraction from a simple preference for encoding from scratch, and (if so), what really motivates that preference. There may be merit to this preference – for instance, correcting OMR may be a 'less musical' task that does not offer the same (or equivalent) pedagogical benefits as transcription-from-scratch. In short, even a sophisticated infrastructure for coordination should perhaps accommodate the option of doing this task ab initio, 'the hard way.'

Correcting OMR is one matter; reviewing each other's work is another. In academic corpus creation this too is mostly conducted behind closed doors, with no release until the final product is ready. This is an understandable position for contexts unaccustomed to sharing work in progress, though error-checking is one task for which the 'many heads are better than one' approach of crowdsourcing is ideally suited. A change of focus from discrete, disconnected products to many 'micro-contributions' model may help here too by diminishing the expectation of perfection prior to release and thus encouraging all-comers to contribute what they have, proof-read and polished or otherwise. Quite apart from the extensive and variable quality of what is hosted online, who can say what treasure troves may be hiding on hard drives around the world, ready to swell the ranks of the world's collection of FAIR and open encoded scores?

Moving Forward Together

While there remain many open questions over the best technical and inter-personal methods for corpus building, some considerations are abundantly clear. First, we are strongly incentivized to work together, so that we can build meaningfully on each other's work and produce large, open corpora for all. Second, we can and should find ways to involve members of as broad a range of musical communities as possible not only in the use but also the creation of corpora. To that effect, organizers should be sensitive to the kinds of tasks that people actually want to do, even if this does not appear to optimize efficiency. Many people around the world happily transcribing from scratch is better than a few people miserably correcting OMR. Third, and most important, we need widespread support for the philosophy of open science extolled in this special edition of EMR: in short, we should release score encodings under the maximally permissive licenses (CC0, PD, or equivalent). Any formal restrictions are a serious barrier to progress.

This is no small matter. Open licenses involve a real commitment to corpus-building as a collective rather than individual effort. To take the thorniest issues head on, consider the 'non-commercial' and 'attribution' options (to use Creative Commons' language). 'Non-commercial' may sound virtuous, but while score encoders opting for this license may imagine themselves fending off a monolithic, exploitative commercial entity, it is much more likely to cast doubt on well-meaning use cases like performance of the score at a charity fund-raiser concert. Even if that specific use is (or would have been) permitted, the lack of clarity means would-be users waste time and energy wondering and will quite likely decide not to risk it.

'Attribution', in turn, is something we should all strive for: we certainly should acknowledge people who put time and energy into projects for the common good where that is practical and important, but here too the strict formality is contrary to broad openness. Wouldn't we do better to commit informally to acknowledging one another where that is important and practical (such as the full disclosure of datasets used in a scholarly study) but not require it formally in all cases (like that amateur concert)? The answer depends in part on our attitudes to engagement with wider musical communities, our openness to all.

As a step towards a more meaningful and productively coordinated effort, it would be helpful to agree on a minimal 'manifesto' or value statement in which we:

• recognize the value and importance of digital scores for musicians, teachers, and scholars to adapt for various purposes,
• likewise, recognize the considerable effort it can take to produce these scores,
• informally undertake to attribute that work where it is practical and meaningful to do so,
• nevertheless consider intellectual property claims inappropriate for the direct transcription (without significant, new editorial intervention) of works in the public domain,
• accordingly believe that the formal license for such transcriptions should be maximally permissive (CC0, PD, or equivalent) and commit to this in our own work.

Open Music Theory Textbook and Anthology

While many textbooks now include a parallel online provision, they generally remain limited in several important ways. First, the number of examples still tends to be rather meagre, despite the fact that page constraints of the printed medium no longer apply. Second, those examples are still very often static images that offer little versatility or functionality beyond their print equivalents. Third is persistence of prohibitive cost. There is a great public appetite for music theory (as attested by vigorous discussion on social media sites and YouTube channels, for instance), but only a tiny percentage of those interested parties would be prepared to pay $100 or more for a textbook where some version of the same content can be accessed for free online. These considerations are part of the motivation for Open Music Theory (Gotham et al., 2021): a textbook, anthology and workbook project in which we have sought to provide free, high-quality material that is both comparable to the traditional print textbook but also benefits from the additional reach and functionality of the online medium.

Among our hypothetical colleagues, Prof. X would be the one most likely to produce or use the 'When in Rome' corpus, and Dr Y would be more oriented towards the textbook, but here too we stand to benefit from thinking across those boundaries. For instance, one of the most fundamental tasks of both a textbook-anthology and a corpus study of analyses is to explore specific instances of a certain chord or progression. This may be in the service of other matters, but even the simple retrieval of examples is of value in itself. Harmonic analysis datasets like 'When in Rome' enable a relatively easy retrieval of this kind of information and can be used to produce musical anthologies of examples at scale. 19

I have included new code in the 'When in Rome' repository for building such an anthology, 20 and implemented one version of this as the 'Harmony Anthology' in Open Music Theory (https://viva.pressbooks.pub/openmusictheory/chapter/anthology-harmony/). This anthology provides lists of examples by topic (standard anthology fare such as augmented sixth chords), automatically extracted from human analyses of CC0 scores in the Lieder Corpus. It bears repeating that while the processing is automatic, the analyses themselves are by people. Among other benefits, the scale of the provision stands to help students contextualize assertions made about so-called 'normal' resolutions. While we may still opt to highlight a few examples in the first instance (as the textbook does), partnering them with links to hundreds of examples – very much including borderline cases – provides a much stronger sense of how exemplary those cases really are of wider practice. It will cost nothing to scale this anthology up as score and analysis corpora grow in the future.

Further, when the dataset of harmonic analyses eventually includes multiple analytical readings of the same works for a larger number of pieces, 21 we will be able to highlight the areas where analysts disagree. These divergences provide important insights both in relation to the particular work at hand, and of the underlying theoretical foundation of the analyst's approach in a more abstract sense. In the simplest case, finding these divergences is readily automatable as the difference between two files. 22 More sophisticated versions can point to the musical nature of the difference between two sources, a subject to which we now turn.

Openness to all: 'The Roman Umpire' code and 'Working in Harmony' app

Scores and analyses hosted publicly under open licenses and amenable formats are valuable, but if we wish to help people develop their skills in tasks like Roman numeral analysis, then we need to offer more. Here, too, routine computational operations developed in the service of research projects can be readily repurposed for these broader uses and may serve, in turn, to grow the corpora.

'The Roman Umpire' is a new code module providing basic feedback on harmonic analyses. For instance, one feedback option involves comparing the pitches implied by each harmony with those in the score at the corresponding moment and alerting the user to moments where the pitch content and/or bass notes diverge. The choice remains with the analyst, the code simply points out moments they may wish to review. I initially developed this code for proof-reading during corpus building; as it proved so useful for experts checking their work, it stands to be doubly useful in pedagogical contexts with a greater need for such guidance. Furthermore, as the process of generating feedback is completely automatic, we can readily offer this as a service that anyone can use as much or little as they chose, whenever they like, via a simple web app.

We have created an initial, alpha version of such an app called 'Working in Harmony' at https://fourscoreandmore.org/working-in-harmony/, 23 initially offering feedback on a user's analysis of any score from the Lieder Corpus. This will prospectively serve to help students enrolled in classes proofread their work more effectively, freeing up class time and instructor feedback for more interesting, musical matters. At least as importantly, it will offer some provision, however rudimentary, to those who wish to learn Roman numeral analysis but have no such class enrollment.

This too can contribute to the creation of FAIR and open datasets and the specific question of involving diverse contributors. We offer this service for free to anyone and invite users to return value as citizen scientists by submitting their analyses (anonymously, unless they choose otherwise) into a corpus which may prove useful to computational musicologists, not least in learning more systematically about the kinds of errors and improvements that students of Roman numeral analysis make in the course of their work. This provides a potential for symbiosis, in which simple code developed in the service of research projects can serve the wider musical public, who in turn help to build up corpora for a range of research projects including those with a focus on developing public-facing pedagogy. 24

Openness to adjustment

Simple though these computational operations may be in principle, there are always procedural questions and challenges to address in the design of such a system which have a direct bearing of the prospective adoption by musical communities, and thus to whom they are truly open. The final parts of this paper consider how computational formats and systems can both promote – but also constrain – openness.

Just as encoded scores (like musicXML) are preferable over 'fixed' formats (PDF) on the basis of their FAIR attributes, so too is that true of analyses: among other tasks, we need to be able to access the constituent information of an analysis in a structured way and to adjust that work in place. For simpler musical passages, analysts often agree about their preferred reading of a passage. That being the case, one of the main benefits of sharing analyses in easily edited formats is the opportunity for other analysts to add a variant reading either within the same file or by making a copy and adjusting only the passages (often few and far between) that they see differently. Editable files speed up the process of producing those analyses by avoiding duplicated work.

In cases of more sustained disagreement, editing an existing analysis in place may not be helpful, or at least not without a more sophisticated user interface. This is not only a question of the number, but also the types of changes. For instance, editing an existing analysis arguably presupposes a method by which the analyst works through from start to finish in a linear fashion, but many analysts do not work this way, preferring instead to start at a kind of middle-ground level, seeking only the main moments indicative of the three harmonic functions (tonic, pre-dominant, dominant) in preparation for a cadence. 25

Engaging with this use case would probably require some kind of multi-layered analysis for harmonic motion at different 'levels', perhaps vertically stacked (in a quasi-Schenkerian style) and aligned with the score, thus enabling users to weave freely between levels. The system could perhaps offer an option to auto-complete certain parts, for instance by extracting the 'big picture' readings from detailed harmonic analyses using a system such as that proposed by Rohrmeier (2011, 2020). This could potentially provide a valuable framework for building students' metacognitive understanding of harmonic analysis. That said, the potential for confusion is high: for instance, when you adjust a detail in the foreground analysis, should the other levels auto-update? User-interface questions of this kind are among the most important for determining how FAIR and open a resource is, and to whom.

On the relative openness of syntax

Openness to adjustment is a high priority for both analysts and for the user-interface of any public-facing app. At the same time, any computational system requires at least some strict syntactical constraints. Specifically, in this case of Roman numeral analysis, we need symbols to have a single, unambiguous meaning. Implementing this requires some careful consideration and warrants discussion.

First, there is some variation in how Roman numerals are encoded, for instance, in the encoding of inversions by Arabic numerals (V42, I6, …) or Latin-alphabet letters (so V42 becomes Vd and I6 is Ib). While the former is probably more common in musicology today, the latter is still in active use. 26 Nevertheless, there is no internal ambiguity: it is clear within either system which pitches are invoked.

Similarly, even within the figured Roman system, there are different conventions, for instance in handling the sixth and seventh degrees in minor. Should 'viio' refer to a chord on the (raised) leading note, or should it require a leading sharp ('#viio')? There are different approaches here, but still no internal ambiguity. Music21's handling of all common variants has improved recently in tandem with our presentation of the 'Roman Text' standard (Tymoczko et al., 2019) 27 and has developed further since, notably through the integration of Nápoles López and Fuginaga (2020)'s unifying regular expression.

Nevertheless, while most of the headaches of apparent 'ambiguity' in Roman numeral analysis are actually attributable to variant, interoperable styles, genuine ambiguity does still creep in at the edges. Abbreviations clarify where this line falls. Perhaps the most common abbreviations in use in Roman numeral analysis handle the cases of:

1. Augmented 6^th chords: for instance, 'Ger65' stands in directly for #ivo65[b3], compressing a complex symbol into a simple shorthand and identifying a canonical chromatic chord. (This class of abbreviation also helps to side-step some commonly confusing matters such as the 'real root' of the so-called 'French' sixth chord).
2. Neapolitan 6ths: 'N6' replaces 'bII6'. The emphasis here is less on compression and more on highlighting a special status for this particular chordal configuration.
3. The Cadential 6/4: an abbreviation like 'Cad64' likewise serves to side-step ongoing arguments, this time over whether the Roman numeral for this chord 'should be' I64 or V64. This abbreviation can also contribute to a more contextual reading by distinguishing cadential from passing usages of the same chord.
4. Common-tone diminished seventh chords: sometimes afforded the label 'Cto7' explicitly to avoid ascribing tonal function (implied by Roman numerals) to a non-functional diminished seventh that shares at least one pitch with at least one of the proceeding and following chords.

The first three cases here involve simple, direct, one-to-one mapping between the chord and the abbreviation, and thus are readily included. The 'Cto7' label, by contrast, does not unambiguously commit to specific pitches and pitch spellings. Often these pitches and spellings can be deduced from context, but that is not universally reliable and so the label is somewhat problematic from a computational perspective.

That being the case, rather than entering 'Cto7' and later attempting to deduce the pitches from there, we may do better to work the other way around, using unambiguous chordal designations at the point of creating analyses and then deducing the presence of a 'Cto7'. For instance, perhaps the archetypical form of the Cto7 sees it prolong the tonic chord (I-Cto7-I), with the root of that prolonged tonic as the common tone, and the Cto7's pitches equivalent to #iio7 (or another spelling with enharmonically equivalent pitches). Some analysts may balk at using any Roman numeral here due to the apparent harmonic function implied, but at least the chord is clear (with or without context) and it is still easy to extract all-and-only these moments for a study of common-tone progressions. 28 Moreover, if this is at the limits of what functional harmony and Roman numerals can express, then it is perhaps a natural limit for what a computational implementation should seek to support.

Some readers will find this a trivial aside, but analysts committed to using 'Cto7' may not be prepared to do without it and may thus reject any resource that does not include provision for it. In short, while we clearly want to support different users and their preferences, we may have to make concessions of this kind in some cases to support a common framework and/or make difficult procedural decisions in the design phase. That said, nothing prevents flexible users from reintegrating these notations into their non-communal uses of the corpora, especially when the computational means to do so are provided (as in the case of editable formats discussed here). We should aspire to create inclusive, open, FAIR datasets, but as discussed above in relation to scores, we should focus on sharing the most mutually useful resources from which more specialised individual projects can grow. Not everything has to be shared.