Hance will demo its kilobyte-size AI audio processing software at TechCrunch Disrupt 2025 (techcrunch.com)

Norwegian startup Hance will demo its kilobyte‑size AI audio processing software at TechCrunch Disrupt 2025, showcasing models that have been shrunk to just 242 kB and run entirely on‑device with about 10 ms latency. Built by a team with deep audio pedigree and trained on Soundly’s high‑quality sound library (from F1 car roars to volcanic rumble), the models can separate sources, remove noise, echo and reverb, and enhance speech clarity in real time. Early customers and partners include Riedel Communications (F1 radio supplier) and Intel, and Hance says its tiny models are already attractive to live‑broadcast, sports, law enforcement and defense use cases where reliability, low latency and energy efficiency matter. For the AI/ML community this is a compelling example of edge AI for audio: achieving professional‑grade signal processing at kilobyte scale highlights advances in model compression, hardware adaptation and dataset quality. The on‑device approach reduces cloud dependency, latency and power use while improving privacy and reliability for mission‑critical audio. Hance’s work to port models to Intel’s NPUs and other chips also underscores the growing need for cross‑platform optimization in real‑time ML. The startup’s non‑exclusive partnerships and fast R&D focus suggest continued iteration — a useful case study for anyone working on tiny, efficient ML models for real‑world deployment.