Current Projects

What I'm building right now.

ML audio research meeting interactive AV prototyping, documented in depth.

Trent Eriksen

Audio ML, audiovisual installations, practice based research.

Accepted at top conferences: NeurIPS 2026, ISMIR 2026 and ISMIR 2025.

Voice conditioned audiovisual installation

Windchime
Gray Area NeurIPS CCRMA, Center for Computer Research in Music and Acoustics Stanford University
Details

Shown at Gray Area, San Francisco; accepted to NeurIPS 2026 Creative AI, Sydney Paper coauthored with Celeste Betancur, CCRMA, Stanford University.

Languages PythonTypeScriptJavaScriptp5.jsThree.jsStrudel (pattern DSL)HTML/CSSShell
Core skills Audio language models (CLAP, CLaMP 3) Speech recognition (faster-whisper) Vector retrieval (FAISS) Live coded audio (Strudel) Generative visuals (p5.js, Three.js) Hardware control (monome) Offline, unattended deployment
May 2026 to present
Explore →
Neural audio instruments for a screenless pedal

HRNSXTN x RDMSXN
MUTEK, international festival of digital creativity and electronic music Elk Audio Roland Future Design Lab Music Hackspace ISMIR 2026, Abu Dhabi ISMIR, International Society for Music Information Retrieval
Details

Music Hackspace × MUTEK hackathon: 1st, Elk Audio Challenge; 2nd, Roland Future Design Lab; presented at MUTEK Forum, Montréal 2026; ISMIR 2026 Late-Breaking Demo, Abu Dhabi

Languages C++PythonTypeScriptShell
Core skills Knowledge distillation Neural audio codecs (music2latent) Embedded realtime C++ (JUCE) On device inference (RTNeural) ARM NEON int8 search Cross compilation (Docker, 32 bit ARM) Granular synthesis Evaluation design
August 2026 to present
Explore →
Live audiovisual assistant for Ableton

Lichtspiel
AIMS 2026, Berklee AI Music Summit Music Hackspace Berklee College of Music
Details

Built at the AIMS 2026 hackathon, Music Hackspace × Berklee AI Music Summit, Boston, June 2026

Languages TypeScriptJavaScriptp5.jsPythonMax (Max for Live patching)HTML/CSS
Core skills Ableton Live integration (Max for Live) Realtime WebSocket bridge (Node) Generative visuals (p5.js) LLM authoring behind validation gates Audio analysis (CLAP, librosa) Hardware control (monome)
May to June 2026
Explore →
Transfer learning for drum audio style classification

Probing the World for Groove
ISMIR 2025, Daejeon KAIST ISMIR, International Society for Music Information Retrieval Universiteit Leiden LIACS, Leiden Institute of Advanced Computer Science Kunstinstituut Melly
Details

ISMIR 2025 Late-Breaking Demo, KAIST, Daejeon; MSc thesis, LIACS, Leiden University, defended at Kunstinstituut Melly, Rotterdam

Languages Python
Core skills PyTorch PyTorch Lightning CUDA Cloud GPU training (A100, L4) Transformer modification (PaSST heads) CNN architecture design Transfer learning Audio augmentation
Thesis June 2025; ISMIR September 2025
Explore →

How they relate

Shared lineage, different problems.

Lichtspiel started from Windchime's visual runtime; HRNSXTN carries the same discipline onto embedded hardware.

Compare them side by side 6 dimensions
Windchime HRNSXTN x RDMSXN Lichtspiel
Type Voice driven installation, runs unattended. Neural instruments on an embedded pedal. Live instrument, played on stage.
Origin Built on Live Muse; Gray Area, SF; NeurIPS 2026. Two podiums at MUTEK, Montréal; ISMIR 2026. AIMS 2026 hackathon at Berklee, Boston.
AI approach Retrieval. Reflects recorded sound back. Distillation: big model offline, tiny live. Generation. Authors new visual scenes.
AI at runtime In the loop, guarded, with failover. 10 Hz worker, never the audio thread. Authoring only, off the render path.
Coupling Standalone: a browser and a monome. Elk Stomp, under its own engine. Inside Ableton Live and Max for Live.
Success looks like A full exhibition day, no crashes. A pedal set, no laptop on stage. A clean three to four minute set.

◈ What Lichtspiel inherits

  • Windchime's p5.js instance mode host loop and render pipeline.
  • The serialosc monome bridge and LED event protocol.
  • Musical meaning mapped to output, not just amplitude.
  • All three fence the model behind validation.
  • All three keep the same decision, release and incident records.