LatentScore at SIGGRAPH 2026
Editable music from natural language.
- Talk
- Sunday 19 July, 9:20 AM
- Room
- 403 A
Published as "LatentScore: Sketching Soundscapes with LLM-Distilled Retrieval for Procedural Synthesis," SIGGRAPH 2026 Talks, ACM Digital Library.
Text prompt in, layered soundscape out, running live on CPU.
- Read the paper dl.acm.org/doi/10.1145/3799818.3812120
- Get the code github.com/prabal-rje/latentscore
- Project site latentscore.com
- Slides (PDF) 25 pages, 13 MB
pip install latentscore What it does
Most procedural audio tools either play canned samples or wait for a beefy GPU to spin up a diffusion model. LatentScore does neither. Text in, structured synth configurations out, runs on a laptop CPU. The synthesizer is the renderer; the LLM-distilled retrieval is what gives it a sense of taste.
I think the pipeline transfers well outside of music too - any time you need responsive, content-aware audio at scale and a sample library is too rigid to be useful.
There's also a companion short paper at NIME 2026 in London.
What else I build
- AI safety. Response-safety classification research in collaboration with the MLCommons AI Risk & Reliability (AIRR) working group.
- Civic data. Lighthouse (in progress), an entity graph over Canada's public records, for journalists and researchers.
- Open-source tooling. SmartPipe, semantic pipes and queries for your terminal; run PDFs, images, audio, video, and text through Unix verbs that understand their input.
- Production ML. Writing on LLM control at prabal.ca.
Say hello
- Email: prabal@rjeinc.ca
- LinkedIn: prabal1997
- Site: prabal.ca