Nick Gellatly
From a walk to a play
Earlier this week I visited the Sedgwick Museum to check the route and now, I have made it talk, as you will see from previous posts on this blog, I had trialled this a week or two ago using ElevenLabs.
The recce on 19th October changed more than I expected. The Whewell Gallery turned out to be a fair walk from the front door, and studying the museum’s own labels gave me some interesting facts I hadn’t known before and the trail ended not at Jonathan Cutbill but in the gift shop, where a Pride T. rex and a Trans Pride T. rex were sitting on the shelf as if they had been waiting for me. So, the first job was a new script, Version 3: nine stops, a long walk through the twentieth century, and a finale among the rainbow badges.
The second job was the big one. Version 3 In the Beginning Was the Word existed only on paper. I wanted it to be something a visitor could scan, press play on and walk with. This post is the story of how it got there, including the bits that went wrong.
Finding the voices
I made the voices with ElevenLabs, using its newest model, Eleven v4. The appeal of v4 for drama is that you can direct it inside the text itself. Square-bracket tags such as [sassy], [whispering] or [long pause] work like stage directions, and the model performs them rather than reading them out.
Before recording anything I rewrote the whole script as a clip list: one row per line, with the character’s name and every direction stripped out, the tags added, and a note of which ear each line belongs in. Version 3 became 167 clips.
Then came casting. I had already designed a voice for each character, but a test recording of Stop 2 showed two problems. Norman Flint, the porter, was fine but a touch too quick for a man who “walks at the speed of continental drift”, so his lines got more pauses and a [very slowly] tag. Crystal Formation was the bigger worry. She is a drag queen made of minerals who bursts out of the amethyst case, and the first voice sounded like a polite tour guide. I designed three new options and chose the third, which has the exuberance and timing the part needs. The difference between “You’re boring them to sediment” said nicely and said properly is the difference between a guide and a character.
Recording 167 clips
I recorded Stops 1 and 2 first and listened before going further, so any problems would show up early and cheaply. They sounded right, so the rest followed: one take per line, spread across four ElevenLabs Flows so each section was easy to find. The whole play used about 29,000 credits.
A few things needed care:
- Names. Text-to-speech guesses at unusual words, and guesses badly. I wrote the hard ones phonetically into the script: Whewell (“Hew-ell”), Nopcsa, Onllwyn and the Dulais Valley, and Jonathan Cutbill’s Latin collection name, Haud Nominandum.
- Lines spoken together. The two rival novelists’ “And then…” and the moment where the left and right ears finally agree on “Jonathan Cutbill. Geologist. Bookseller. Out gay man.” were recorded as separate clips and laid over each other later.
- One fix. Listening back, I changed Darwin’s “pounding the NW mountains” to “pounding the North Wales mountains”, so a listener hears it clearly, and re-recorded that one clip.
In total the play runs to about 32 minutes of speech before any sound effects or music.
The label and the life
The central idea of the play is that museum labels and real lives often disagree. The Catalogue reads the official version (“Resided with her cousin by marriage”) and Crystal tells you what it leaves out. I wanted listeners to feel that, not just hear it, so the play is built for stereo earbuds:
- Left ear: the Catalogue, the voice of the label.
- Right ear: the voices of the lives themselves, such as Anna Gurney, Darwin’s notebook and Nopcsa.
- Centre: Crystal, Norman and the Clerk, who guide you through.
The payoff comes at Jonathan Cutbill, the last stop. After a whole tour of the two ears disagreeing, both say the same words at the same moment, and for once the label and the life match.
To make editing manageable, I assembled the clips into eleven voice tracks, one per stop, already in order and already panned. Each came with a label file for Audacity, which marks every line by its clip number and every gap with the sound that belongs there, so opening a stop shows a ready-made map of what goes where.
Sound effects and a song from 1917
The script needs 44 sound effects, from keys jangling at the front door to a glitterball spinning in the Darwin gallery and a till drawer in the gift shop. I worked from a checklist, one row per cue, with search words and a column for the file name and credit, and sourced about 50 sounds from three places:
- Freesound.org, for most of the everyday sounds. Many need the creator named, so I kept a credit for every one.
- BBC Sound Effects, for period sounds such as the courtroom and the typewriter, used under the BBC’s licence for educational use.
- Pixabay, for the comic ones, such as the squelch and the gasp, which need no credit.
The music was the real headache. The play ends on Ivor Novello, whose last musical, Gay’s the Word, gave Jonathan Cutbill’s bookshop its name, and I wanted its big number, “Vitality”, running underneath. I could not find a royalty-free recording anywhere. A 1951 West End show that was barely recorded is not much use to an audio play in 2026.
In the script, Crystal mentions that Novello wrote “Keep the Home Fires Burning” at twenty-one and the whole country sang it. I found John McCormack’s 1917 recording, which is old enough to be out of copyright, as is Novello’s tune. It now plays like a distant gramophone through the long walk, growing a little louder each decade, comes up when Crystal introduces Novello, and returns for the finale. A song about keeping a light on until people come home turned out to fit a play about queer history better than the one I had planned.
Mixing, and getting it to visitors
The first full mix placed each sound at its marked gap, faded it in and out, and set it well below the voices, with ambient beds such as the seagulls at Anna Gurney’s stop and the swamp at William Smith’s map sitting very low under the speech. Every stop was then brought to the same loudness, a standard level for listening on phones, so nobody has to reach for the volume between stops.
Not everything arrived. Seven of my sound files were too large to transfer, including the footsteps, the ice cracking and a 139 MB recording of a music-hall piano, so the first mix used substitutes or left those gaps quiet. I listened to every stop twice with headphones and decided it worked as it was. Those sounds would have made a difference but I had to bear in mind that this version is a prototype.
For hosting I chose SoundCloud, first because the museum already uses this platform for its commentaries and second because it plays straight from a phone browser with no app and no sign-in. The play is now a public album, In the Beginning Was the Word, with eleven tracks in walking order and content notes in the titles of Stops 5 and 6. The play can be accessed using this QR code or the link below it.
The last link in the chain is to add the QR code. It will be on page 2 of the A6 zine, but will not be in the printed version as I have already made these arrangements – a bit of cut and paste will be required. It is also on the draft invitation web page beside a new “Listen to the play” button (I have not yet made this page live). A visitor can scan at the door, put their earbuds in and press play.
Reflection
What worked. Testing small before committing big. Recording two stops first, and a single scene before that, meant the voice problems surfaced when they cost a few hundred credits to fix rather than thousands. The clip numbering also earned its keep: one number tied together the script, the ElevenLabs clips, the Audacity labels, the sound checklist and the credits, so a note like “5-01” meant the same thing everywhere.
What surprised me. How much of audio production is file management. Some of my hardest moments this week were not creative at all: a zip file Windows would not open, a cloud folder that could not be shared, upload pages that hid their buttons. Next time I will keep sound files small (MP3 rather than huge WAVs) from the start and keep everything in one place.
On using AI. This project used AI at two levels. The character voices are synthetic, made with ElevenLabs, and I am open about that in the credits. I also worked with Claude, an AI assistant, as a production partner, for rewriting the script into a clip list, assembling and panning the voice tracks, laying in the sound effects and troubleshooting along the way. The decisions stayed mine: the story, the casting choices, which voice Crystal should have, which sounds to use and when the mix was right. I found it most useful for the repetitive, technical work, which left me more time for the parts that needed a human ear. The open question I want to keep thinking about is how synthetic voices sit alongside a project about real, often silenced, people. My answer for now is that the voices are characters, clearly made, while the history they tell is real and sourced.
Next steps. A test walk at the museum with my phone and earbuds, to check that each stop lasts about as long as the walk to the next star, especially the long walk to Cutbill. Finally, I need to work on how best to present the ‘pitch’ on 12th November as there are a few elements to demonstrate.
Leave a Reply