Publishing and Performing Digital Musicology on the Web: Signature Sound Vienna

2.1 Musicologists’ attitudes to digital methods

ISMIR papers have repeatedly identified an understanding of user contexts, information needs and behaviours as a prerequisite to developing useful MIR systems (Lee and Cunningham, 2013; Schedl et al., 2013). Reflecting on the first 10 years of ISMIR (2000–2009), Downie et al. (2009) called for actively encouraging user participation, including by musicologists, to help create ‘truly useful music‑IR systems’ (p. 17). Though digital resources are important to contemporary musicologists (Inskip and Wiering, 2015; Laplante and Sauvé, 2023), scholars are heterogeneous in self‑assessed technological expertise, report struggles in integrating software within their research practices and may even perceive digital approaches as a threat to traditional musicology. Tools supporting traditional methods are more readily accepted than those conducting distant forms of analysis on the scholar’s behalf – though digital capacities for grounding assertions in empirical data are understood to strengthen the validity of findings (Laplante and Sauvé, 2023, p. 37).

Consequently, MIR technologies are rarely used in musicology: Analysing the 1,055 ISMIR papers published during 2012–2021, Borsan et al. (2023) found only 28 reporting on datasets, methodologies, code and/or tools whose use is subsequently reported in the musicology literature, even when surveying interdisciplinary venues such as the International Conference on Digital Libraries for Musicology and the Journal of New Music Research.

2.2 FAIR approaches to music research

Provenance is an important notion in both traditional and digital scholarship, (i) placing evidence objects (e.g. performance recordings; scores) within historical, cultural, social and pragmatic contexts, particularly in terms of ownership and possession, and (ii) modelling scholarly assertions and analytical processes and their outcomes in terms of agency, parameterisation and evidence basis. Correspondingly, contextual information on agents, materials and contexts has been incorporated into musical and musicological data models using the Linked Data paradigm, informed by FAIR (‘Findable, Accessible, Interoperable and Reusable’) principles of research‑data management (Wilkinson et al., 2016). Originating in the natural sciences, these principles have received growing interest in music – see, e.g., Dix et al. (2022), Hofmann et al. (2021) and Weigl et al. (2021) – and more broadly in the digital humanities (e.g., Hawkins, 2022).

The OMRAS2 project (Cannam et al., 2010) pioneered the use of Linked Data approaches to managing music information. This was achieved in part through the creation of a family of Semantic Web vocabularies and ontologies centred on the Music Ontology (Raimond and Sandler, 2012), which has been applied in research but also in broadcast media institutions. The Transforming Musicology (Crawford et al., 2017) and FAST (Sandler et al., 2019) projects respectively investigated the application of such approaches in music scholarship and in music industry contexts. AcousticBrainz (Porter et al., 2015) and CALMA (Bechhofer et al., 2017; Page et al., 2017) processed decentralised, user‑contributed information, publishing large collections of user‑provided recordings and music‑analytical metadata as semantically‑enriched Web corpora. Dunya (Porter et al., 2013; Serra, 2014) established a system for exploring multicultural music research collections, assembled as part of the CompMusic project (Serra, 2014), by means of culturally‑relevant contextual relationships. The TROMPA project investigated the connection of such repositories over the Web, allowing their data to drive task‑specific interfaces catering to diverse user communities, including music performers and enthusiasts alongside scholars (Weigl et al., 2019, 2021). Polifonia took further steps in this direction, establishing new shared ontologies and tools for data integration, transformation and unified access to music heritage collections (de Berardinis et al., 2023). The ongoing LinkedMusic project aims to lower barriers of access to music search, harnessing large language models to provide internationalised natural language querying across a large, heterogeneous collection of music datasets (Pond et al., 2025).

The Music Encoding Initiative’s MEI format (Hankinson et al., 2011) has seen widespread application as a machine‑readable, music‑semantic frame of reference in many of these projects. MEI provides a means of encoding music notation documents, typically rendered as visual scores using the Verovio toolkit (Pugin et al., 2014). MEI documents are readily parsable by machine agents. By producing visualisations in the scalable vector graphics (SVG) format retaining the MEI encoding’s (XML) element hierarchy and identifiers, Verovio facilitates implementing browser‑based applications centred upon interactive digital scores (Pugin, 2018).

MEI’s musically‑meaningful XML model affords granular integration with Linked Data (Weigl and Page, 2017), which may flexibly target music elements using their XML identifiers in media‑fragment uniform resource identifiers (URIs) (Troncy et al., 2012) or via music‑specific selection schemes (Viglianti, 2016). MEI and Linked Data technologies have been combined in a number of research projects connecting scores and music recordings with extra‑musical context, including in modelling annotations and scholarly claims (Lewis et al., 2019, 2022; Sanfilippo et al., 2025), in associating musical meaning with timed media (Lewis et al., 2019) and in dissemination (Dreyfus et al., 2025; Lewis et al., 2018). A complementary approach is taken by Dezrann (Ballester et al., 2025), a platform for the annotation and synchronization of music data, which integrates (rather than interconnects) disparate multimodal datasets by converting them to custom unified data representations.

3.1 Workflow for corpus creation

We have acquired and digitised source materials on physical media (CDs, DVDs and LPs) relevant to the NYC series (Weigl et al., 2022). Their acquisition is non‑trivial: only NYC recordings from 1987 onward remain available for purchase. We thus searched second‑hand shops and flea markets to expand our corpus, incorporating other orchestras’ recordings of frequently‑performed NYC compositions by consulting Bielefelder Katalog Klassik and other catalogues. Our collection includes more than 80 albums comprising audio from 61 New Year’s Concerts (43 ‘complete’ concert recordings, alongside compilations) and 30 albums by other orchestras performing relevant repertoire.

In digitising our physical media collection of audio recordings, we are informed by the approach taken by Dunya (Porter et al., 2013) to manage large collections of recorded audio and associated metadata. We captured our tracks as uncompressed WAV files. As in Dunya, MusicBrainz is used to store and organise the editorial metadata obtained from covers and booklets of our physical media. Our tracks were identified using the open‑source AcoustID audio‑fingerprinting service¹ and labelled with discographic metadata obtained from MusicBrainz using the Picard Tagger (Stutzbach, 2011). Labels were validated against (physical) liner notes. Corrections or previously missing records were contributed back to AcoustID and MusicBrainz where required. Validated metadata were converted to readily‑processable format using Picard’s ‘Generate Cuesheet’ plugin (one cuesheet per recording), including structured information on both album and track levels. A custom cueToRdf Python script processed these sheets, extracted audio features and generated Linked Data descriptors to characterise our collection. The script generates RDF graphs adhering to the Music Ontology, with distinct sub‑graphs listing predicates and objects for all subjects (entities) of the following types:

• mo:Record, a music record (CD, LP, DVD, etc);

• mo:Track, a track on a record;

• mo:Signal, the audio signal captured on a track;

• mo:Release, a record as a commercial product (e.g. with liner notes and a catalogue number); and

• mo:Performance, a (recorded) musical performance by one or more agents or groups of agents (e.g. a conductor, an orchestra, a soloist).

An instance‑level graph is illustrated in Figure 1.

Figure 1

Entity URIs may be dereferenced to access their corresponding data in our corpus. MusicBrainz URIs are provided where relevant.

In Dunya, digitised and semantically enriched music data are stored within a set of relational and graph databases. To minimise technical requirements and promote sustainability and workflow reusability beyond our project scope, we instead store our Linked Data graphs as static files (Section 3.2). An additional union graph was produced to facilitate cross‑dataset operations (e.g. SPARQL queries) and to provide a convenient representation for archival in versioned research data repositories such as Zenodo.

Serialisations are provided in different RDF formats, allowing machine‑agents to specify their preferred content type for processing (Section 3.2.1). Audio‑derived features are published following a non‑consumptive approach to reproducible research on copyright‑restricted materials (Organisciak and Downie, 2021). Waveform envelopes are computed to provide accessible visualisations of recordings for public dissemination (Section 4.4). Industry authority metadata (e.g. International Standard Recording Code) are provided to facilitate copyright‑conformant access by project‑external researchers.

3.2 Publishing our Linked Data corpus

w3id.org is a URL redirection service operated by the W3C Permanent Identifier Community Group providing identifier namespaces – in our case, https://w3id.org/ssv/ – for the formulation of proxy URIs pointing to a resource’s hosted location. Content‑negotiation allows machine agents to specify their preferred RDF serialisation format through standards‑compliant specification of acceptable content type(s). This service facilitates interoperability and supports long‑term sustainability by allowing the dataset to be moved at a future date without impacting (potentially project‑external) Linked Data referencing its resources, following established best practices for implementing FAIR Linked Data on the Web (Garijo and Poveda‑Villalón, 2020).

Our data is stored on GitHub. Hosting is provided free of charge, with expectations for long‑term storage. Cross‑origin resource sharing compliance permits Web applications to directly load and process our resources. All files are versioned, with access flexibly provided through predictable ‘raw’ GitHub user‑content URIs. The data can thus be exposed in a specified state according to publication context by incorporating a label corresponding to a particular branch, tag, release or commit number. Versioned data can be explicitly exposed using w3id.org proxy URIs.

4.1 Mei‑Friend

Music scores enjoy a privileged position as evidence objects in scholarship and provide a natural reference when comparing collections of recorded renditions. MEI encodings are highly suitable for such purposes (Section 2.2), but manually encoding MEI‑XML is laborious and requires significant technical expertise. We have developed mei‑friend⁴ (Figure 2) to accommodate: a browser‑based music encoding editor (Goebl and Weigl, 2024) capable of importing and converting music encodings in various formats to MEI using an integrated instance of the Verovio toolkit. Pre‑existing score data obtained through OMR or typeset using MuseScore thus bootstrap the encoding process before conversion to MEI for the ‘last mile’ of correction, validation and enrichment. An adaptable user interface presents a synoptic view of both MEI XML and (Verovio‑generated) graphical scores. Notation elements selected in the visual score are correspondingly selected within the XML editor and vice versa. Graphical menus with keyboard shortcuts support modification or insertion of MEI elements. An integrated GitHub module facilitates collaborative encoding. A facsimile panel permits direct comparison and interlinking with (regions of) scanned source documents. A MIDI player sonifies the score, highlighting notes and following the score during playback. An enrichment panel supports editorial interventions and markup, including creation and display of annotations using the MEI <annot> element. Stand‑off Linked Data structures adhering to the Web Annotation Data Model (Sanderson et al., 2017) and the Music Annotation Ontology (Lewis et al., 2022) may be generated or imported, supporting independent annotation of Web‑hosted MEI files. These annotations are Linked Data structures addressable from external contexts, including other Web applications, for subsequent enrichment and reuse.

Figure 2

4.2 Listen Here!

Listen Here! is a Web application⁵ for machine‑assisted close‑listening to aligned audio collections (Weigl et al., 2023). Recordings are listed in a side‑panel and are loaded into or removed from a comparison view on user request. Dozens of recordings may be simultaneously visualised as playable waveforms using WaveSurfer.js,⁶ and load‑on‑demand access is provided for arbitrarily large numbers of recordings. Tempo‑curve visualisations are available as waveform overlays, showing either absolute performance tempo (quarter notes per minute) or relative deviation from the average tempo at a given moment across configurable subsets of the collection.

Audio visualisations are vertically stacked (Figure 3), recalling the interface of the Sonic Lineup desktop application.⁷ On playback, the sounding (active) waveform is highlighted; a cursor indicates playback position and can be manually reset via mouse‑click. By clicking on another waveform, it is made active and the previous one becomes inactive; playback seamlessly switches to the new recording, using the alignment grid projections to arrive at the corresponding playback position with fine‑grained accuracy. Visual indicators may be set at corresponding time positions across exemplars. Optionally, playback jumps back to the closest marked position when switching between recordings, greatly facilitating close listening to specified moments in the music across renditions.

Figure 3

Hand in hand with mei‑friend’s support for stand‑off annotations, we have implemented equivalent functionalities for audio recordings within the Listen Here! tool. Linked Data structures targeting MEI elements may be imported and projected onto corresponding (audio‑to‑score‑aligned) timeline positions in each rendition. Annotations are visualised through corresponding coloured regions upon each loaded waveform. Using a Select Recordings button, the Linked Data are enriched with timed media‑fragment URIs referencing timeline intervals on the selected recordings.

4.3 Data‑Level integration

Our tools support decentralised workflows in which Linked Data structures are generated and iteratively enriched using independent Web applications targeting different music modalities. This is achieved by adherence to explicit machine‑interpretable vocabularies and data models – an inherent affordance of the Linked Data paradigm – casting the processing of music information as ‘digital music objects’ experiencing ‘semantic music flows’, accreting data and contextual information along the way (De Roure et al., 2018; Sandler et al., 2019).

Our applications implement this approach using the Solid platform for social Linked Data (Mansour et al., 2016), a decentralised solution supporting contribution and sharing of user‑created content, including for digital music research applications (Weigl et al., 2020). Solid Pods are both identity providers (allowing users to authenticate when logging in to independent applications) and user‑administrated storage spaces supporting selective data sharing across applications and with other users.

The Music Annotation Ontology (MAO; Lewis et al., 2022) provides a framework to capture, extend and refine cross‑modal music annotations. Both mei‑friend and Listen Here! are capable of generating, loading and enriching MAO Linked Data structures (e.g. by adding new MAO Selections to existing Music Objects). Unlike in Sanfilippo et al. (2025), the primary aim is not to model the granular semantic structure of musicological claims but rather to map annotations to their targeted music representations, which may be interrelated across different modalities and varying levels of abstraction.

To facilitate interactions with this model, we apply a storage and retrieval pattern to aid discovery of the relevant data across applications. A user logs in to different Web applications (currently, mei‑friend and Listen Here!, though other applications, e.g. presenting music‑analytical visualisations, would be equally supported) using Solid. On authentication, each application establishes the existence of a music‑annotation data–integration container within the user’s storage, linked from the user’s Solid profile. Similarly, containers for MAO music objects, for Web Annotations pertaining to music data integration and for a resource discovery service are established.

Each application presents Web resources – in our case, MEI encodings and performance recordings – to the user, who is given the option to make a selection upon the presented resources and to identify this selection as the basis of a new music object. This prompts the creation of a chain of Linked Data structures within the MAO music object container, starting with a MAO Selection object that directly targets the selected resource fragments using media‑fragment URIs – here identifying MEI elements or time intervals on audio recordings. The process is repeated with increasing abstraction for MAO Extract and MusicalMaterial – see Lewis et al. (2022). The discovery service container is checked for existing collections referencing a particular Web‑resource (e.g. MEI file, audio recording). These collections are modelled using schema.org, a widely‑used vocabulary for the description of Web resources, as DataCatalogs⁸; each collection is about⁹ its respective Web resource. Where such a collection does not yet exist, it is created, and references to each MAO music object or Web Annotation structure are added in connection to the resource being annotated, using the dataset¹⁰ property.

This pattern enables independent applications to establish and interact with shared storage locations for mutually‑relevant data resources. Code‑level integration is not required; the interoperability required to enact the ‘music flows’ across applications as described here is achieved through integration on the data. Figure 4 provides a simplified, idealised illustration of such a workflow using our tooling.

Figure 4

4.4 Disseminating our findings through hypermedia publication

Using our approach, scholarly insight is explicitly connected to constituent evidence objects (score fragments; audio‑recording intervals), allowing findings to be disseminated as hypermedia publications juxtaposing scholarly analysis, annotated digital scores and audiovisual materials. Using Linked Data to model both evidence and scholarly annotations allows each of these entities to be addressed from external contexts, including from more traditional forms of dissemination such as journal articles – see, for example, our long‑form report of an investigation into Johann Strauss II’s Kaiser‑Walzer (VanderHart and Weigl, 2025). In short, this approach supports decentralised scholarly communication, annotation, re‑interpretation and re‑contextualisation in the spirit of Linked Research (Capadisli, 2020).

Music scholars cannot be assumed to possess expertise in Linked Data models, serialisations or other particular technologies, nor can such expertise be required if data adhering to these models are to be made useful to this music‑domain‑expert audience. Appropriately user‑friendly views over the relevant data structures are required.

To realize such human‑readable hypermedia publications, we have developed Primal¹¹ (Figure 5), a browser‑based viewing and interaction tool for music annotation Linked Data. The tool is instantiated with the URI of a Linked Data entity (supplied via the ?obj query parameter); configurably traverses along graph connections originating from this ‘root’ entity, building an internal graph representation as it proceeds; and registers encountered target objects of interest for custom rendering. In its default configuration, the tool assumes the root entity and graph relationships relevant for traversal to adhere to the Web Annotation or MAO data models. Target objects are specified using media fragment URIs of MEI elements and specified intervals of audio recordings, in compliance with the Music Encoding and Linked Data framework (Lewis et al., 2018).

Figure 5

Following traversal and subsequent validation, specialised renderers construct human‑readable views of different aspects of the data. These include:

• descriptive content of scholarly annotations, rendered as text;

• annotated score fragment(s), rendered to SVG using Verovio, with annotated elements highlighted;

• annotated audio interval(s), highlighted upon corresponding waveform visualisations^12;

• metadata, including the identity of the annotator and information on the annotated recordings;

• A navigable graph visualisation showing a configurable subset of the currently‑traversed Linked Data graph, generated using the open‑source Mermaid.js diagramming tool^13; and

• A full data listing of the graph contained at the currently‑navigated RDF resource (highlighted on the visualisation), presented as nicely formatted JSON‑LD.

The different views are presented in sections, navigable using a persistent menu. Sections are intentionally ordered to be welcoming to all music‑interested users, starting with textual, score and audio contents and relegating the more technical metadata, graph visualisation and data listing (JSON‑LD) views. These latter views remain readily discoverable and accessible, allowing users with greater technical affinity, musicologists interested in explicitly engaging with the underlying data and students in pedagogical contexts to gain deeper insight into the underlying corpus and data model.

5.1 The process

SSV operationalises a hybrid methodology combining close and distant listening. Our tools facilitate iterative cycles of exploratory listening, structured annotation and Linked Data–driven analysis across a multimodal corpus of historical and contemporary performances.

Listen Here! enables comparative close listening across aligned audio‑to‑score collections, facilitating the identification of both fine‑grained performance nuances and larger structural differences within and across recordings. We used this tool to closely examine the 10 most frequently performed works in the New Year’s Concert repertoire – all composed by members of the Strauss dynasty (e.g. An der schönen blauen Donau, Radetzky‑Marsch, Kaiserwalzer, Pizzicato‑Polka, and the Fledermaus overture). Multiple renditions of each work were accessed in rapid succession through time‑aligned playback anchored to MEI‑encoded scores. Corresponding historical (public domain) scores were encoded and annotated using mei‑friend, enabling the interactive selection of notated elements, the creation of Linked Data annotation structures with persistent URIs and their reintegration into Listen Here! for targeted analysis and enrichment. This annotation supports a layered analytical process. Listening sessions began with open‑ended, hypothesis‑generating questions – e.g. whether the Vienna Philharmonic exhibits a consistent interpretive profile across decades, or the extent of conductors influence on interpretive contours. Initial subjective observations were subsequently formalised as semantically‑enriched annotations, which served both as analytical anchors and as citable evidence.

Annotations and linked media were contextualized using the Primal platform, which renders the resulting knowledge graph as human‑readable hypermedia. These interconnected evidence objects, accessible via URIs, support transparent scholarly communication, data reuse and further inquiry beyond the scope of the project. This integrated workflow – anchored in traditional close listening but enhanced and rescaled through bespoke computational tools – allowed the project to move beyond anecdotal claims and toward empirically grounded and replicable analysis of interpretive trends, institutional identity, performance style and conductor agency.

5.2 Findings

We conducted close comparative listening across several dozen recordings spanning more than 80 years of New Year’s Concerts and related performances. This analysis yielded insights into evolving performance practices, institutional habits and interpretive norms within the repertoire of the Vienna Philharmonic and beyond. The following, diverse examples illustrate representative analytical outcomes.