From Page to Place: Reviving Forgotten Hymns through AI in a Site-Specific Heritage Music Project

From Page to Place: reviving forgotten hymns through AI in a site-specific heritage music project Ross M. Purves 1[0000-0003-4301-7024] 1 Department of Culture, Communication and Media, University College London Institute of Education, 20 Bedford Way, London, WC1H OAL [email protected] Abstract. This paper documents a site-specific heritage music project that revived three hymns from The Home Hymn Book (1885) for the 170th anniversary of Foster Hill Road Cemetery in Bedford, England. Drawing on the cemetery’s rich musical and civic history, the project aimed to reanimate forgotten local repertoire through a blend of live vocal recordings and AI-generated choral textures. Central to the process was ai-choir, a free, open-source tool developed by Ralf Popescu, which implements the So-VITS-SVC deep learning framework for singing voice conversion. The resulting performances were designed to emulate the heterophonic, reverberant sound of a Victorian church congregation and were presented within the cemetery’s restored Gothic chapel using its audiovisual system. The project highlights the creative potential of AI in heritage music contexts, particularly where resource limitations preclude live ensemble performance. It also raises important ethical questions about the role of AI in community musicmaking, especially in relation to amateur participation and the social value of local musical engagement. While the AI-generated voices introduced artefacts and limitations, careful mixing and spatialisation techniques helped produce convincing and evocative results. The project demonstrates how AI tools can support the rediscovery and reinterpretation of place-based musical traditions, offering new ways to connect communities with their cultural past. It also suggests avenues for further research in choral acoustics, psychoacoustics, and the aesthetics of AI-mediated performance. The final recordings are available at: https://youtu.be/tZr3mHwlMr4. Keywords: Local cultural resonance, hybrid orchestration techniques featuring AI, Questions of musical ethics, authenticity and artistic integrity. 1 Introduction This paper presents a site-specific heritage music project that revived three Victorian-era hymns from The Home Hymn Book (1885) to mark the 170th anniversary of Foster Hill Road Cemetery in Bedford, England. Deeply rooted in the town’s cultural and musical history, the project sought to reanimate forgotten local repertoire through Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 883 2 R. Purves a blend of live vocal recordings and AI-generated choral textures. Central to this process was the use of ai-choir, a tool developed by Ralf Popescu that leverages the SoVITS-SVC (Soft Voice Conversion using VITS) deep learning framework for singing voice conversion. 2 The heritage music project Foster Hill Road Cemetery in Bedford (formerly known as Bedford Cemetery) was opened on 5th June 1855, forming the town’s first municipal place of rest [1]. Covering some 37 acres, the site accommodated around 100,000 burials until a new cemetery was established in 1987. Some 16,000 memorials are to be found at Foster Hill, and those buried there include many prominent figures in Victorian and Edwardian Bedford along with others from around the world. The cemetery features a large chapel building in Gothic revival style (Fig. 1a). Fig. 1. (a) Bedford Cemetery Chapel [credit: Dennis Simpson, CC BY-SA 2.0, via Wikimedia Commons]; (b) The DVD showcase reel running via the chapel’s AV system during pre-event tests. Stereo sound was produced from the speakers built into both screens in reasonably good quality. Officially now a ‘closed cemetery’ with very few burials still taking place, much of the ongoing care of the site is undertaken by a group of volunteers known as the ‘Friends of Bedford Cemetery’. This charity’s aims are to ‘promote and preserve the environmental, cultural and historical benefits of [the] Victorian Cemetery and Chapel complex’ [2]. In 2016 the Friends successfully bid for funds from the Heritage Lottery Fund to convert the old chapel building into a visitor and resource centre, enabling the local community and those from further afield to be appropriately welcomed and informed of the site’s historical importance [3]. The refurbished chapel hosts the Friends’ annual programme of public talks and it was through presenting research at one such event that my ongoing collaboration with Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 884 From Page to Place 3 the group began [4]. In June 2025, Foster Hill Road celebrated its 170th anniversary and I was asked to provide music for a weekend of commemorative events [5]. Those buried at the cemetery include several important musical figures from Victorian and Edwardian Bedford, amongst them published composers. Whilst their sacred and secular repertoire was popular and admired in its day, the vast majority is now long forgotten. So was born the idea of seeking out scores in the archive and reviving a representative range through specially-recorded performances, aiming to provide a poignant and site-specific musical soundtrack to the weekend’s commemorative activities within the chapel. Minimal time and no financial resources were available for this project so it was essential that all recorded performances could be realised using my own practical musical and technological skills, with only very targeted support from fellow musicians. In total, six musicians’ and lyricists’ works were represented in the resulting ‘showcase reel’, embracing late Romantic piano pieces, musical comedy, parlour songs and hymns [6]. This ran from a 25-minute looped DVD, routed through the chapel’s recently-installed audio visual system (Fig. 1b). Screens of information and images accompanied each piece. It is the recording and production of the three hymns included within this showreel which forms the focus of this paper. 2.1 The hymn repertoire Hester Periam Hawkins (1846–1928) is a reasonably well-known figure in Bedford’s history and beyond due to her pioneering work as an astronomer, together with various social and philanthropic endeavours. She is also remembered as the wife of Joshua Hawkins, a five-time Mayor of Bedford who laid out much of the town’s historic green and built environment [7]. Hester and Joshua now lie together at Foster Hill Road. Whilst recognised in academic hymnology literature [8], Hester’s work as a hymn composer, lyricist and hymnal editor is much less known locally. Her first hymnal, The Home Hymn Book of 1885 was compiled with collaborator Edwin Moss (1838–1919). Subtitled ‘A manual of sacred song for the family circle’, it was designed for domestic use and reflected the Victorian ideal of family worship and musical engagement at home. Hester’s ability to mobilise Bedford’s musical community of the time appears to have been central to the book’s success; it includes contributions from several Bedfordassociated composers and writers, many of whom lie not far from her at Foster Hill Road Cemetery. Some of the book’s melodies bear names of Bedfordshire villages, further reinforcing the local character of the work. For the purposes of the showreel project, the first edition copy of the Home Hymn Book held in the British Library was consulted, and a representative set of three hymns selected for realisation. These were as follows: • Oakley (‘Sun of my Soul!’): the melody of this hymn is attributed in the book to ‘R Rose’, understood to be Robert Rose (1814–1898), who set words by John Keble (1792-1866). Robert Rose was a prominent Bedford-based music teacher, organist and music shop owner who now lies at Foster Hill Road. The name ‘Oakley’ seems Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 885 4 R. Purves likely to refer to a nearby Bedfordshire village, although further research is needed to confirm the nature of this link. • St Pancras (‘Hours, and Days, and Months, and Years’): this is a setting of words by J.S.B. Monsell (1811–1875) by Robert Rose’s son Henry (1854-1911). Henry maintained Bedford music teaching connections whilst also serving as organist at St Pancras Parish Church in London. He is also now interred at Foster Hill Road. • Sunnyside (‘Oh, happy home! where Thou art loved the dearest’). This features a melody by Hester herself, setting words by Karl Johann Philipp Spitta (1801–1859) as translated by Jane Borthwick (1813-1897). This melody was named after the house on Linden Road, Bedford, where Hester and Joshua raised their large family. The text’s themes of familial love and children make this a touching choice. 2.2 Digital engraving and live recording The four-part organ accompaniment for each hymn was first digitally engraved into MuseScore Studio (musescore.org) and then exported both as PDF documents (for reference during recording) and MIDI files (for importing into digital audio workstation (DAW) software). The hymn melodies from within these four-part textures (represented by the scores’ top lines) were then sung into the DAW a male singer (the author) in both tenor and bass registers (i.e. an octave part). A first female singer then recorded the melodies in both alto (in unison with the tenor) and soprano registers. A second female singer then added a further soprano recording of each melody. 3 Applying ai-choir In an ideal world, time and resources would have been available to facilitate the recording of a live choir performing each hymn. However, this was not the case. To this end, it was decided to create the sound of a full church congregation singing these hymns using singing voice conversion software. Although commercial tools such as Synthesizer V Studio Pro (dreamtonics.com), Vocaloid (vocaloid.com) and ACE Studio (acestudio.ai) exist and are well established in the music production field, these were not viable for the present project and a free, open source alternative was sought. The tool ai-choir was released by musician, producer and technologist Ralf Popescu in 2024 [9]. It provides a simple Python-based command line ‘wrapper’ implementation of the So-VITS-SVC framework [10] for the end user. Invoking the tool and with the name of a wav file containing an exemplar monophonic sung vocal will result in a folder of generated files including four ‘female’ voices, three ‘male’ voices, together with a stereo, blended version to which convolution reverb has been applied. A range of configuration options is available via an accompanying json file. In the present case, only the individually-generated voice files were used and not the blended, summed output. The tenor and alto recordings performed by the live singers were selected as source audio files for separate input into ai-choir. Given that the end goal was to convey the sound of a church congregation, rather than a trained church choir, it was important to Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 886 From Page to Place 5 use more than one source audio file so that the resulting, combined AI voices produced an appropriately heterophonic texture i.e. one not perfectly synchronised in terms of rhythm and phrasing. Using two audio files ensured that there were subtle timing and phrasing differences between the AI voices produced from each. Moreover, the tenor and alto parts were chosen as these were both sung in the same, mid-range vocal register. Since the voice models provided with ai-choir included those emphasising soprano and bass timbres, using mid-range source performances ensured that the AI parts were produced in musically convincing human vocal ranges. 3.1 The So-VITS-SVC pipeline Ai-choir makes use of the So-VITS-SVC Stable 4.1 fork [11] of an earlier open source project published by the SvcDevelopTeam [12]. The original ‘Soft Voice Conversion using VITS’ deep learning-based framework (known as ‘So-VITS-SVC’) was developed to support the computer sound and music academic and hobbyist communities, particularly anime enthusiasts desiring to have their character creations 'sing' during animation. So-VITS-SVC departs from the text-to-speech basis of the forgoing ‘Variational Inference Text-to-Speech’ (VITS) approach, adopting instead a process of 'singing voice conversion' featuring an integrated pipeline of four main stages. Within the default implementation used by ai-choir, these main stages are as follows. 1. Linguistic and phonetic content from the source audio file is extracted using the content encoder ContentVec [13]. 2. The musical pitch contour perceived by human listeners (F0) is extracted from the source audio using Parselmouth [14]. 3. The outcomes of stages 1 and 2 are fed into a pre-trained neural network stored within a PyTorch Model Checkpoint (.pth) file) [10]. Seven So-VITS-SVC .pth files are available within ai-choir, representing the four female, three male voice models. The tool cycles through each voice model in turn, using the extracted qualities of the source audio as the basis of an output ‘inference’ based on the pre-training. 4. Audio output is generated from each model inference using the neural vocoder NSFHiFiGAN [15]. After completing this pipeline, ai-choir continues by mixing the individual audio files and applying convolution reverb. However, since these stages were not employed in the current project, they are not described in detail here. 4 Blending the live and AI generated voices with organ in the final mix Once the 14 AI voice parts had been generated for each hymn (i.e. seven from each source performance), these were imported into DAW software alongside the organ part and original live voice recordings. It was observed that the AI voices featured a very Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 887 6 R. Purves small but noticeable, consistent latency in comparison to the live voices and so it was necessary to offset their starting positions forward very slightly to compensate. Given the desire noted above to ensure a heterophonic texture consistent with an untrained but enthusiastic and ‘lusty’ Victorian church congregation, the AI voices based on the different source audio files were interspersed and panned across the stereo image to emulate the sense of space within a church nave (Fig. 2). Fig. 2. An indicative stereo image of the various human and AI voices When listened to individually, it was notable that some of the AI voices featured noticeable digital noise artefacts or raspy qualities. However, the perceptibility of these issues was masked to a reasonable extent when all AI voices were blended. Additionally, it was felt likely that age and health variations within a real Victorian choir would have probably resulted in an equally diverse range of vocal timbres. Nonetheless, all AI voice tracks were routed to a submix where considerable artificial reverb was applied (with an RT60 value of 1.5s, emulating decay times found in the smaller English parish churches surveyed by Mapp [16]). The output balance was set to emphasise the ‘wet’ rather than ‘dry’ output (i.e. the reverberated sound) to help ‘smooth over’ the various artefacts and rasps whilst also helping emulate the sound of a large choral ‘wash’ experienced within a traditional English parish church due to the strong sound reflections from the stone walls. Whilst this significant use of reverb was effective in improving the sound of the AI generated voices and helping to emulate appropriate church acoustics, it also had the disadvantage of blurring the overall melodic lines of the hymns and reducing the Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 888 From Page to Place 7 intelligibility. Therefore, the live audio vocal performances were not submixed in the same way but instead subjected to a separate reverb process which maintained the rt60 value but emphasised the dry sound with only minimal ‘wet’ output. These live voices were also set at slightly higher volumes than the AI voices. The psychoacoustic net result was that these live voices were perceived in the foreground of the mix, adding human presence, emphasising the contours the hymn melodies, and providing important additional clarity and definition to the sung words. A further step was to pan the live vocal parts to sit within the overall stereo image created with the AI voices (fig. 2). Only moderate panning was applied in most cases, however, as it was unknown at the time whether the AV playback system in the cemetery chapel was capable of full stereo playback or whether, for instance, only a single channel would be routed to speakers. With regards to the organ part, minimal humanisation was undertaken on the imported notation file (specifically, reducing note lengths by around 5% to emulate an organist lifting their hands and re-shaping for the next chord). Although it had originally been hoped to render the organ part using a SoundFont sample set of a suitable English parish church organ, this was not possible in the time available so an organ patch was used from a software synth bundled with the DAW. The organ remained central in the stereo image for all three hymn realisations and also had reverb applied. 5 Aesthetic and ethical reflections Overall, the final mixes for these hymns, as played within the cemetery chapel, were deemed to be convincing and satisfying from the perspective of musical aesthetics. The prominent positioning of the human voices in the mix helped give the impression of an overall ‘live’ church congregation. This was further enhanced by the use of two different source audio files to generate the AI voices and the interspersing of these across the stereo image. Careful application of artificial reverb also helped not only to enhance the AI voices and their blending, but also to emulate a typical English parish church acoustic. The hymn realisations were well received by those attending the cemetery anniversary celebrations, and there were also questions about how and why AI voices were used (I had made no secret of the process employed, citing the ai-choir tool on the accompanying text slides and provided greater detail in a handout). The realisation process, comments and questions received led to many broader ethical reflections on my part. Bedford has many amateur choirs [17] and it could be argued that an opportunity was missed to involve them within this hymn realisation project. Much has been written over the years regarding fears that sound synthesis and, latterly, AI might displace professional performing musicians and impact on their livelihoods [18]. In this case, the unfunded nature of the project meant that there was never any prospect of being able to pay musicians. However, this project offers an important reminder that computer-based sound synthesis and AI processes may also impact the opportunities and corresponding intangible benefits for amateur musicians as well. Not only might projects such as that described reduce performance and ensemble experience opportunities for local amateur singers and musicians but they might also Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 889 8 R. Purves negate the social, communal and place-based benefits of amateur music making—not to mention opportunities in this case to engage with the cultural history of one’s hometown. Given what research now confirms regarding the personal and social benefits of musical engagement [19] and also the rise in the use of musical activities as part of social prescribing within healthcare [20], these are potentially serious considerations which those in the computer music community might find it helpful to keep under review. A further ethical dimension relates to the original purposes of the realised works as Christian hymns. All three composers evinced strong and demonstrable commitments to the Christian church, and it was important that the realisations of these hymns not only did justice to their original sacred purpose but were also heard in an appropriately sensitive space (a (still) consecrated chapel in the immediate vicinity of these musicians’ final resting places). Beyond this context, however, this project raises further religious considerations. Although not directly employed in a worship context, the AI voices were nonetheless still ‘singing’ words which would have been intended by the composers and lyricists to be heard by God. This resonates with emerging literature which considers the human and religious authenticity of involving AI within aspects of religious worship [e.g. 22]. Moreover, and developing the points above regarding the potential disenfranchisement of local amateur singers, there are also questions regarding missed opportunities for local religious communities to be involved in what might have been a meaningful act of site-specific worship. Such perspectives link the wider, ongoing debates over the possible marginalising of amateur performers and congregations in inclusive worship music [e.g. 23]. On the other hand, in the present case it can also be argued that the reviving of these hymns under significant time and financial constraints resulted in effective awarenessraising ‘showreel’ versions of sufficient quality to potentially stimulate interest in such repertoire and perhaps lead to re-adoption by local church congregations and choirs. Moreover, their inclusion within a high-profile commemorative weekend, attended by both cross-section of the local population and individuals with significant local political, civic and social influence, could help revive the memories of important historical residents of Bedford and their work. Here again, then, the computer music community might also usefully reflect on the extent to which their activities could help shed light on forgotten or unfamiliar musicians and their repertoire, and bring these to life for the wider community through innovative and impactful technologically-mediated presentation. 6 Conclusions and implications An advantage of a tool like ai-choir is that it runs locally and, beyond the initial downloading of the .pth models (which can be very time consuming, given their size), no cloud services are required. This is useful in allaying concerns regarding control over intellectual property and data protection. Although the tool is time intensive (with around 1–1.5 hours of processing time required to produce all the output audio files for a 2–3-minute hymn recording), a further benefit is that it does not require energyand Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 890 From Page to Place 9 water-intensive cloud-based server farms to operate [21]. Its simple interface and open source status are also significant, potentially helping to maximise critical engagement with AI frameworks and approaches in ways accessible to relatively non-specialist musicians and those on limited budgets. In turn, this might lead to even more informed perspectives on ethical debates such as those highlighted above. On musical and technical levels, this project suggests significant potential for future research and practice-based exploration in regard to psychoacoustics and perception of phrasing in AI-generated voices, along with choral acoustics and how reverb affects perception of ensemble singing within English parish church settings. As noted, the organ sound and stereo positioning was not particularly authentic and greater realism might also have been possible if all reverb processing was undertaken using convolution impulses recorded in an English parish church themselves. Experimenting with alternative So-VITS-SVC voices models (which ai-choir supports) might also have resulted in higher quality voice outputs overall. Nonetheless, this project enabled the creation of a full choral sound without the logistical and financial burden of assembling a live choir. This helps confirm that tools such as ai-choir, together with the underlying framework used, are useful for producing demonstration recordings of forgotten repertoire. Whilst not yet entirely viable for professional recordings the results are certainly promising for mock-ups and educational use. Moreover, the project sparked interest and debate about the future of choirs and AI in music and raised awareness of such matters amongst the cemetery anniversary weekend audience. The final hymn recordings can be heard at: https://youtu.be/tZr3mHwlMr4 Acknowledgments. I am grateful to the Friends of Bedford Cemetery for the invitation to participate in this project and for their support during its realisation and this write-up. I also wish to thank my fellow singers, Helen and Mel Purves. Thanks are also due to Nicolas Gold, for his excellent feedback on a draft version of this paper. Disclosure of Interests. I am a self-funding member of the Friends of Bedford Cemetery but received no financial support to complete this project. References 1. The History of Foster Hill Road Cemetery, https://fosterhillroadcemetery.co.uk/historyfoster-hill-road-cemetery, last accessed 2025/06/13. 2. Charity Commission website for ‘The Friends of Bedford Cemetery (Foster Hill Road)’, https://register-of-charities.charitycommission.gov.uk/en/charity-search/-/charity-details/4019562, last accessed 2025/06/13. 3. Heritage Lottery Fund: Schedule of decisions under delegated powers to Head of HLF East of England on 23 February 2016, https://www.heritagefund.org.uk/sites/default/files/media/corporate/eeddecisions201602.pdf, last accessed 2025/06/13. 4. Bedford’s Musical Diemer Family, https://fosterhillroadcemetery.co.uk/bedfords-musicaldiemer-family, last accessed 2025/06/13. 5. 170th Anniversary Event, https://www.facebook.com/FosterHillRoadCemetery, last accessed 2025/06/13. Proc. of the 17th International Symposium on CMMR, London, UK, Nov. 3-7, 2025 891