Date: 14/10-28/11/2024
Project Showcase
Project file
Please download to a PC from here.
Downloading instruction
- You will need a VR headset that can be connected to the PC that you will use to view the project.
- Download the original file.
- Download the Epic Games Launcher. Launch the software and download Unreal Engine version 5.3.
- Open Unreal Engine, go to Browse, find the project file and click Open.
- Click the Play button on the toolbar (usually in the middle of the top of the screen, looks like a small triangle ▶️).
- Navigate the game with physical movement, and interact with the model with hand gestures (hand gestures are shown below)
Hand gestures demonstration for game interaction (also shown in the video below)
- Use a finger of your left hand to touch the menu to access other maps.
- Use your left hand to grab the glowing object to interact.
- Pinch the right-hand index finger and thumb to aim and teleport/move.
If you cannot open it or do not have a VR headset, please view the video demonstration below.
Video demonstration (Second half of the video)
If you would like to try it out yourself but do not have suitable equipment, please contact the researcher, Yuan Gao, from the “Contact Me” section in the navigation bar at the top of the screen. Yuan will arrange an in-person showcase with you at the University of Leeds (only available before 30th October 2025).
AI-generated accented voice-over
A comparison of the outcomes with subtly adjusted settings
Reflective Diary
Following the first version of the Dalton Mill VR project, participant feedback and critical reflections from conferences hosted by the Science Museum Group catalysed the development of a second, upgraded iteration. This version shifted focus from basic environmental reconstruction to interactive and narratively enriched experiences, emphasising user engagement, AI integration, and enhanced fidelity in both sound and model design.
My role in this stage was centred on three key contributions: (1) audio and video editing, (2) upgrading the loom’s 3D model, and (3) designing an AI-based text-to-speech (TTS) voice for narration in Yorkshire dialect.
The visual model upgrade involved rescanning the moquette loom by dividing it into six smaller sections using photogrammetry, followed by model cleaning and mesh merging in Blender. LiDAR remained unreliable, particularly for cylindrical and reflective components. I hand-measured the supplement parts of the loom (e.g. levers and spool rack), and recreated it manually in Fusion360 with the help of my husband, Gregory Harris. After software version mismatches between collaborators were resolved, these two modelling formats were eventually merged and reassembled in Unreal Engine.
The second major task involved creating a Yorkshire-accented AI narrator. We planned to use this narrator to guide audiences through the VR experience to provide a sense of the location of Dalton Mills’ historical background. Firstly, I used ChatGPT to generate the working process of Mill workers based on the interview we conducted with a volunteer at the Calderdale Industrial Museum and information available online. The AI was asked to create this working process as a senior worker teaching a new worker (which is the audience) to use the moquette loom. Then, working with Greg again, as he is from West Yorkshire, I collected his voice for training samples. Interestingly, after testing over 15 AI voice TTS models (from HuggingFace and TTS platforms via Google search), I could not successfully generate it. I tried to figure out why. By chance, I went to The Great Yorkshire Shop at the Corn Exchange in Leeds, and realised the Yorkshire accent can be written in a different way, which directly shows the pronunciation of the accent. Therefore, I used a Yorkshire accent translator by Mr Dialect (https://mr-dialect.com/en-gb/translator/english-yorkshire/) to translate the script from ChatGPT to Yorkshire-accented text. Surprisingly, it worked, and even though it is not effective on all 15 TTS models, Eleven Labs showed the best outcome so far, which was confirmed by Greg, the voice contributor. However, this AI-generated voice-over was not used in this version due to the time limitation, but an initial version of the AI TTS was used in version 1 (room 3), also demonstrated in the video above.
This upgraded version was showcased at the Manchester Science Museum Group conference, the University of Leeds XR Symposium, and the Developing Healthy Engagement Network Conference (DHENC). During the latter two conferences shown above, I demonstrated to attendees the AI-generated Yorkshire-accented voice-over. It only took two guesses to figure out where the accent is from in the University of Leeds XR Symposium, and visitors’ feedback highlighted the narrator’s voice as “authentic and grounding” and praised the improved machine model as “visually and sonically convincing”. Most importantly, the sense of “working soundscape” was retained and deepened. However, in DHENC, even though it only took one guess for audiences to figure out where the accent is from, a participant from Yorkshire (who attended the practice 10 as well) offered a very critical piece of feedback that “for someone who doesn’t live in Yorkshire, this voice might sound like a Yorkshire accent to them. But for someone like me, who has lived in Yorkshire for a long time, this accent sounds a bit strange, and besides, Yorkshire has many different accents, so for me, this accent doesn’t have a strong sense of regional specificity”.
This comment was a revelation for me, and it also raised more questions. Firstly, is the understanding of an unfamiliar culture based on an individual participant’s stereotypes of that culture? Can this also be extended to how participants understand non-vocal soundscapes? (Which I carried on exploring in practice 10.) Secondly, although technology can acoustically approximate a dialect, an ambiguity of belonging follows: to whom does this “copied” voice actually belong? Is it the AI, the project team, or the community that originally carries the dialect? This controversy has parallels with the issue of cultural appropriation in the context of Indigenous museums, suggesting that digital replication can inadvertently infringe upon a community’s cultural sovereignty. Due to the time limitation, I could not investigate deeper into my research, but it could be explored further by other researchers.
Reflective Methodological Note
Dalton Mill Version 2 marked a methodological transformation from experimental reconstruction to experiential narration. I am no longer presenting a site, but reconstructing a way of working and speaking, embedded in space, dialect, and industrial rhythm.
This reframing led me to experiment with a modular architecture for VR design. It breaks the experience into flexible segments that could accommodate multiple narrative paths, user roles, and sound conditions, and allows participants to choose which parts they are interested in experiencing first. Each interaction is no longer isolated, but a node within the experience that can be rearranged, reinterpreted, or expanded. In this sense, the site became a platform, not a replica.
That modular thinking also affected how my team and I handled the collaboration. Our software incompatibilities between Unreal Engine versions, between Blender and Fusion360 meshes, revealed hidden assumptions in our workflows. But rather than seeing these as setbacks, I treated them as reflection points: why did we prioritise certain render qualities over others? How could we better document our pipelines for future iterations? In this way, friction became a methodology in itself, which is a means of surfacing values, intentions, and blind spots within our creative process. I also learned how technological friction can enhance practice. The challenge of synchronising file formats, versions, and pipelines forced us to communicate better, document our process more carefully, and refine our creative priorities. Far from hindering progress, these frictions grounded our collaboration in mutual translation between disciplines, such as sound design, modelling, software, and storytelling. Combining the cooperation in this series of practices with practices 5 and 6, I proposed a multi-stage collaborative approach for small and medium-sized museums in my research essay Chapter 5.4.2, which can be further tested and iterated in the real museum scenarios and related research.
This shift made me acutely aware of the accent as heritage. The TTS development process was both frustrating and illuminating, which reminded me that AI doesn’t just struggle with language mechanics, but it also lacks social context. An accent is more than sound; moreover, it is a carrier of memory, class, and belonging. Our inability to replicate that with AI reinforced my idea that digital heritage should remain open to human mediation.
Ultimately, this version of the project proved that sound and interaction can turn digital heritage into multi-vocal memory spaces that allow machines to speak, workers to whisper, and code to become/form conversation.
