> ## Content Index > Fetch the complete content index at: https://www.thedaringcreatives.com/llms.txt > Use this file to discover other available public pages before exploring further. # SONAMB Trains Her Own AI on Field Recordings From Manchester's Cotton Mills - URL: https://www.thedaringcreatives.com/creator-stories/sonamb-cotton-mill-neural-synthesis/ - Published: 2026-10-12T13:00:00.000Z - Updated: 2026-10-12T13:00:00.000Z - Description: Vicky Clarke, who releases music as SONAMB (@sonamb__), trains a small neural network on sounds she recorded herself, from the cotton looms at Quarry Bank Mill to glass and electricity, then composes with what the model gives back. Her own notes cover every step. - Author: William Smith - Tags: Creator Stories, artist profile, sound art, music, field recording, neural synthesis, machine learning, musique concrete, installation, #medium:Music, #method:Mixed, #handle:@sonamb__, #tool:PRiSM SampleRNN, #tool:StyleGAN, #tool:Ableton, #tool:Blender, #issue-2026-10 Vicky Clarke knows where every sound in her AI model came from, because she recorded all of them. The Manchester sound artist, who releases music as SONAMB, has been building small datasets out of her own field recordings since 2019: cotton-mill machinery, canal water, glass fragments, metal and electricity. She trains an early neural synthesis model on them, listens to what comes out, and composes with it. She's worth studying because she writes the whole process down. [Her own site](https://vickyclarke.org/?ref=thedaringcreatives.com) walks through the datasets, the training settings, how she listens back, and the moment the model produced a sound she says she couldn't have made herself. You can follow her as [@sonamb\_\_ on Instagram](https://www.instagram.com/sonamb%5F%5F/?ref=thedaringcreatives.com) and hear the records on [SONAMB's Bandcamp page](https://sonamb.bandcamp.com/?ref=thedaringcreatives.com). Her newest piece, Latent Spaces, took all of that and put it in a basement in a Manchester textile building, so people could walk around inside it. ## Recording the dataset herself The method took shape during a residency at NOVARS that began in 2020, the University of Manchester's centre for innovation in sound, working with PRiSM at the Royal Northern College of Music. She framed it around one research question, which [her residency write-up](https://vickyclarke.org/2022/10/11/aura-machine-research-site%5F-machine-learning-musique-concrete-residency%5Fnovars%5Funi-of-manchester/?ref=thedaringcreatives.com) still leads with: "how can concrete materials and neural networks project future sonic realities?" Her background is musique concrète, the tradition of recording real-world sounds and treating them as raw material, so the first step was recording. [For the first piece, Aura Machine](https://vickyclarke.org/2021/08/13/aura-machine/?ref=thedaringcreatives.com), she sorted her field recordings into three classes: "Echoes of Industry" (Manchester mill spaces), "Age of Electricity" (DIY technology, noise and machinery) and "Materiality" (glass fragments and metal sound sculptures). Then she trained PRiSM SampleRNN on them. In a [Q&A with Sound and Music](https://soundandmusic.org/post/vicky-clarke-in-motion-latent-spaces-q-and-a/?ref=thedaringcreatives.com), she explained what that looks like from the inside: > "Each cycle is called an epoch, so on epoch one the output samples are very noisy, as training goes on, the network 'learns better', so in later epochs you start to recognise materials from the input dataset, but reconfigured in an uncanny way." The code came from Dr Chris Melen at PRiSM, who updated it and taught her how to use it. It's [open source on GitHub](https://github.com/rncm-prism/prism-samplernn?ref=thedaringcreatives.com). She has also been clear about why she built her own datasets: "I developed my own methodologies of building concréte datasets of my own recordings, so I knew the origin of every sound, undertook all the labour and the model used relatively small amounts of local computational training at RNCM." You can hear the whole journey on the Aura Machine EP, released on LOL Editions in April 2025\. The middle track is a 20-minute live recording from NOVARS that moves from the original field recordings, through a section representing latent space, into purely generated material. On either side of it she put raw SampleRNN output, at early and later epochs. *"I wanted to share what the source material was for the piece, to show its primitive and noisy architecture."* [Aura Machine by SONAMB](https://sonamb.bandcamp.com/album/aura-machine?ref=thedaringcreatives.com) SONAMB — *Aura Machine* (LOL Editions, 2025). Tracks 1 and 3 are raw SampleRNN output. On Bandcamp. ## "Glass static" Early epochs come out as noise. Further into training, the model started returning sounds that weren't in her recordings, and she named one of them "glass static." She described it to Sound and Music: > "I can hear fragments of noise and glass from the dataset, but it is a new sound that I as a human couldn't create and is quite beautiful. The machine did produce an aura." Vicky Clarke — "First ML sample\_Glass Static," on Vimeo. She kept the model's flaws in the music on purpose. SampleRNN outputs audio at a 16K sample rate, which is lo-fi, and it had what she called "machine wind," which she describes as ghostly, with a presence of its own. She wanted the piece to treat the model *"almost as though the model is an instrument in itself with its own timbre."* That lo-fi character surprised her. On [the Aura Machine Bandcamp page](https://sonamb.bandcamp.com/album/aura-machine?ref=thedaringcreatives.com) she wrote: *"I was enchanted by the timbre of this early SampleRNN model's output material, with its analogue sound, surprisingly reminiscent of early Concrète tape recordings, its lofi character due to a 16K sample rate."* So she followed it back to tape. For Latent Spaces she started learning tape machine techniques so she could process the machine learning material onto tape, which she called *"another transmutation!"* ## The cotton mill Her next project, NEURAL MATERIALS, built a new dataset around one place. [Her notes on the project](https://vickyclarke.org/2024/07/19/neural-materials/?ref=thedaringcreatives.com) describe recording the machinery at Quarry Bank Mill, a cotton mill built in 1784 and now run by the National Trust. She got into the Cotton Processing Room and the Weaving Shed with directional and stereo mics, and recorded the Draw Frame, the Slub Roving machine, the Carding Machine, the Ring Spinning Frame and the Lancashire Loom. The mill is personal. *"My Dad trained as an apprentice electrician here in the 70s, cycling from Hulme to Styal everyday, so it's a special place for me."* The dataset traced a route into the city: cotton machinery on the outskirts, then the canals, then the new apartment blocks in the centre. Each stretch became a data class: cotton, water and noise. After training, she wanted to know what rhythms the model would invent. She put the original machine recordings and the model's exports side by side in Ableton, sorted by epoch and by temperature, the setting that controls how random the predictions get. Converting the output to MIDI drums was the clearest way to hear it. She called the result *"eclectic and brilliant non-sense that a human drummer could not play."* She then sent those pulses as CV triggers into her modular synth. For the live show she built a hybrid rig. Ableton linked to the modular. A Bela PEPPER module ran a PureData patch that played and stretched the AI cotton-mill samples. Two Bela TRILL touch sensors were fitted to one of her metal sculptures, so she could improvise with the samples by touching it. Digital artist Sean Clarke built an audio-reactive system in TouchDesigner from her source images. SONAMB — NEURAL MATERIALS, live, from the [Cyborg Soloists channel on YouTube](https://www.youtube.com/@cyborgsoloists?ref=thedaringcreatives.com). NEURAL MATERIALS debuted at Modern Art Oxford in spring 2024, after a two-year R&D commission from [Cyborg Soloists](https://cyborgsoloists.com/?ref=thedaringcreatives.com), the UKRI-funded programme at Royal Holloway led by Dr Zubin Kanga. She dedicated the piece to her dad, Vince Clarke, who died in July 2024. ## Walking inside the model Latent Spaces, which premiered at Seesaw Space in Manchester in October 2025, is her first solo installation. She made it through Sound and Music's In Motion composer scheme. The idea came out of the training work: latent space is the point in a network where the model works out patterns in its data, and she wanted to know what it would be like to stand in it. > "I'm attracted to this idea of statistical alchemy, that these computational spaces can be mythical and full of potential where sound material is in flux." [The exhibition page on her site](https://vickyclarke.org/2025/09/22/latent-spaces-exhibition/?ref=thedaringcreatives.com) lays out how it's built. The sound is the mill machinery, electricity, glass and metal, run through SampleRNN and spatialised around the room. The visuals came from StyleGAN, trained locally on a custom dataset she calls "Techno-Symbolic." Metal sculptures of her Aura Machine icons, a set of symbols she designed to stand for sound objects in latent space, sit among cotton drapes and moving images. The page also carries a short note on how the technology was used. All content is original material, with no dataset scraping, and training happened locally at RNCM with Dr Christopher Melen. Latent Spaces, a spatial sound installation by Vicky Clarke (SONAMB). Film by Pedro Küster, on Vimeo. For the technical side of the room, she covers routing Ableton for spatial audio, along with field recording and neural synthesis, with Caro C on [the Sound On Sound podcast](https://www.soundonsound.com/people/vicky-clarke-neural-synthesis-sos-podcast?ref=thedaringcreatives.com). Tools Vicky Clarke credits herself - [PRiSM SampleRNN](https://github.com/rncm-prism/prism-samplernn?ref=thedaringcreatives.com) — the neural synthesis model, trained on her own field recordings - [StyleGAN](https://github.com/NVlabs/stylegan?ref=thedaringcreatives.com) — the Latent Spaces visuals, trained on her own symbol dataset - [Ableton](https://www.ableton.com/?ref=thedaringcreatives.com) — analysing the model's rhythms, and the live and spatial setup - [Bela](https://bela.io/?ref=thedaringcreatives.com) PEPPER and TRILL — playing the AI samples by touch, through a PureData patch - [Blender](https://www.blender.org/?ref=thedaringcreatives.com) — 3D models of the metal sound sculptures ## Where the work comes from Before any of this, Clarke was building DIY electronics and sound sculptures, including with the project Noise Orchestra, and she won an Oram Award in 2020\. [Her debut SONAMB album, SLEEPSTATES](https://sonamb.bandcamp.com/album/sleepstates?ref=thedaringcreatives.com), came out in 2022\. It was built partly from transmission recordings between Manchester, Berlin and St Petersburg, and mixes DIY electronics, sound sculpture, musique concrète and machine learning. Teaching runs alongside it. One aim of the NOVARS residency was to demystify machine learning for other artists, and she ran Dreaming with Machines, an education project for young women with Brighter Sound, where participants built their own datasets and audiovisual pieces. In the Sound and Music Q&A she points to artists pushing back on scraped training data, including Holly Herndon's work on public diffusion. We covered Herndon's voice model in [our Holly Herndon profile](https://www.thedaringcreatives.com/creator-stories/holly-herndon-voice-model-holly-plus/). The model family links her to [the Dadabots neural death metal project](https://www.thedaringcreatives.com/creator-stories/dadabots-neural-network-death-metal/) too, which trained SampleRNN on band recordings to generate whole albums. She's also clear-eyed about where her model sits now. *"The technology has moved on so fast in five years, that now the model and piece feels almost retro compared to present generative models."* Latent Spaces still runs on it. [Her post on the Latent Spaces film](https://vickyclarke.org/2026/02/12/latent-spaces-film/?ref=thedaringcreatives.com) links a touring document for the installation, and her [Vimeo channel](https://vimeo.com/user59853476?ref=thedaringcreatives.com) has years of test clips, from the first glass static sample to the Quarry Bank Mill recordings.