Underdog Unveils 32MB Woof Speech Model for On-Device Audio Generation
Artificial intelligence startup Underdog has showcased a compact speech generation model called Woof that operates entirely locally within web browsers without requiring an active internet connection. Measuring only 32MB in file size and comprising 31 million parameters, the open-weight audio model demonstrates that conversational voice synthesis can run directly on consumer phones and personal computers. To demonstrate the technology, the company released an interactive retro-themed browser game featuring a simulated room containing a CRT monitor, desk lamp, and virtual typewriter. Users can type arbitrary text into the typewriter interface to hear it spoken back instantly, while an interactive tuner generates distinct instrumental musical styles in real time.
Retro Browser Demonstration and Parameter Efficiency
The demo opens with a satirical in-game advisory warning players about intrusive cloud providers, underscoring the startup's philosophical opposition to centralized server architectures. Underdog emphasized that the 31-million-parameter model is open for public download, targeting decentralized edge AI applications that prioritize data privacy. While compact models of this scale cannot match the comprehensive capabilities of massive frontier neural networks, they provide sufficient fidelity to power local voice assistants and personalized offline audio feeds. The browser demonstration follows Underdog's recent release of Bark, a 167MB speech-to-text model designed for offline device-level transcription and dictation. The Woof speech model is compressed into an exact storage footprint of 31.7 megabytes, according to Emergent.
Local Edge AI Architecture and Founder Vision
Underdog founder Sigil Wen told TechCrunch that he created the software ecosystem because he sought an independent alternative to cloud note-taking services like Granola. The founder expressed conviction that mainstream consumer hardware will be capable of running state-of-the-art open models locally within months. In addition to its lightweight audio models, Underdog released a 27-billion-parameter language model engineered to run locally on consumer laptops. Founder Sigil Wen highlighted a corporate focus on American-designed, open-source technology granting people total ownership over their personal information, according to Emergent.
Offline Assistant Ecosystem and Edge Deployment
The startup has also opened a waitlist for a forthcoming application positioned as an entirely locally hosted alternative to proprietary assistants like Muse and Instinct. The releases reflect an expanding technical movement toward edge computing aimed at reducing server infrastructure costs and eliminating latency bottlenecks. By packaging capable voice models into a compact 32MB file size, the startup demonstrates practical pathways for private on-device intelligence. Industry researchers noted that while giant tech firms race to train increasingly massive models requiring immense datacenter power, local models reduce compute demands substantially.
Background
Publishing the creation under open-source terms grants programmers full freedom to pull the repository and embed it into custom software, according to Emergent.