Underdog shows off a 32MB speech model in new demo

Why it matters

Underdog, a startup working on small AI models, has released a browser-based game showcasing Woof, its new speech generation model designed to run on everyday devices without an internet connection. The game lets users type text on a virtual typewriter that is spoken back to them, and generate music with a tuner, all processed locally in the browser. The model is openly downloadable and has a very small footprint compared with typical large models. The release follows Bark, Underdog's earlier speech-to-text model for edge devices aimed at transcription and dictation. Founder Sigil Wen told TechCrunch he built the technology because he wanted users to own their data rather than rely on cloud-based notetaking tools. He said he expects capable models to run on devices people already own in the near future, and frames the effort around open, private, locally hosted AI. The company has also released a laptop-capable model and is building a locally hosted app positioned as an alternative to existing assistants. The story illustrates a counter-trend to the industry's push toward ever-larger models requiring heavy processing power.