Underdog Woof: 32MB Speech Model Runs Entirely in Browser

Underdog has officially launched Woof, a 32-megabyte text-to-speech model that runs entirely within web browsers without requiring cloud connectivity. The startup released an interactive browser game on November 6, 2026, to demonstrate the model's capabilities, featuring a retro-themed office environment where players can type text and hear it spoken back instantly through locally-generated voice synthesis.

Interactive Game Demonstrates Local Voice Synthesis

The demo game presents users with a vintage office setup containing a typewriter, CRT monitor, and functioning desk lamp. Players can type anything on the virtual typewriter using their keyboard, and the Woof model generates speech output in real-time without sending data to external servers. The game also includes a music tuner that generates different audio styles through rotation controls.

According to Underdog's announcement, the game serves as both entertainment and proof-of-concept for private, device-based AI interactions. The intro screen warns players about "intrusive" cloud providers, positioning the demo as a critique of cloud-dependent AI architectures that transmit user data off-device for processing.

Technical Specifications and Model Architecture

Woof contains 31 million parameters compressed into just 31.7 megabytes of storage space. The model runs on most modern devices without internet connectivity, executing speech synthesis calculations locally in the browser environment. Underdog has released the model as open-source software, allowing developers to download and integrate it into their own applications.

While Woof's compact size limits its capabilities compared to larger cloud-based speech models, the startup suggests it provides sufficient quality for local assistant experiences, personalized podcast generation, and privacy-focused voice interfaces. The model's browser compatibility eliminates installation requirements, enabling developers to deploy voice features through standard web technologies.

Release Date and Product Context

Officially launched on November 6, 2026, Woof joins Underdog's expanding portfolio of edge-optimized AI models. The company released Bark, a 167-megabyte speech-to-text model for transcription and dictation, just weeks before the Woof announcement. Underdog has also developed a 27-billion-parameter language model designed to run on laptop hardware.

Founder Sigil Wen told TechCrunch that Underdog's mission centers on enabling capable AI models to run on consumer devices within months, emphasizing American-made, open-source development that gives users data ownership. The company is building a waitlisted application positioned as a locally-hosted alternative to cloud note-taking services, reflecting Wen's desire to create personal AI systems that don't depend on external servers.

Edge AI Strategy and Market Positioning

Underdog's approach contrasts sharply with the industry trend toward larger, more computationally intensive models requiring significant processing power. The startup targets use cases where privacy, offline functionality, or low latency outweigh the performance advantages of cloud-based alternatives.

• No internet connection required for voice synthesis operations
• Complete data privacy through local-only processing
• Zero-latency voice generation without network round trips
• No usage costs or API rate limits for developers
• Works on devices with limited connectivity or in restricted environments

The open-source release enables developers to customize and optimize Woof for specific applications without vendor lock-in or subscription dependencies. Underdog positions this transparency as both a technical advantage and philosophical commitment to user autonomy.

What This Means for Browser-Based AI

Woof demonstrates that meaningful AI capabilities can fit within the constraints of browser-based applications, expanding the design space for privacy-preserving voice interfaces. As edge hardware continues improving and model compression techniques advance, locally-executed AI features may become standard components of web applications rather than premium cloud services. Underdog's dual release of speech-to-text and text-to-speech models suggests a broader strategy to provide complete voice interaction stacks that operate entirely on user devices, challenging the assumption that capable AI requires centralized infrastructure.