AI startup Underdog has released Woof, a 32MB text-to-speech model designed to run locally on smartphones, laptops, and other devices without relying on cloud computing. The company introduced an interactive browser game to demonstrate the model’s capabilities, allowing users to generate speech and experiment with sound directly on their devices.
The model contains 31 million parameters and is available for public download. Unlike larger AI systems that require substantial computing infrastructure, Woof is designed for devices with limited processing power and memory.
Woof Demonstrates Local AI Speech Generation
Underdog has introduced an interactive demonstration game featuring a retro-themed room with a typewriter, CRT monitor, desk lamp, and music tuner. Users can interact with the objects to experience the model’s speech and audio generation capabilities.
Typing text on the virtual typewriter causes Woof to read the content aloud, while adjusting the music tuner generates different musical sounds. The demonstration runs entirely within a web browser using the company’s compact speech generation model.
According to Underdog, Woof can operate on most modern devices without an active internet connection once the necessary software is available locally. Its small file size could make it suitable for applications such as personal voice assistants, offline narration, and personalized audio content.
The company has made the model available for developers to download and experiment with, allowing them to incorporate local speech generation into their own applications.
Underdog Expands Its Collection of On-Device AI Models
Woof follows the recent introduction of Bark, a 167MB speech-to-text model developed for offline transcription and dictation. Together, the two models support converting spoken language into text and generating speech from written content.
Underdog has also released a 27-billion-parameter AI model designed to run on laptops. The company is developing a locally hosted AI assistant positioned as an alternative to cloud-based consumer agents, with access currently managed through a waitlist.
Founder Sigil Wen said he started building the technology after wanting a personal note-taking system that could operate on his own devices rather than depending on external AI services. His goal is to give users greater control over their information while making capable AI models accessible through hardware they already own.
Smaller AI Models Target Privacy and Offline Applications
Underdog’s approach focuses on reducing the processing requirements of AI models so they can operate directly on consumer devices. Local processing can reduce reliance on internet connections and allow certain tasks to be performed without transmitting user inputs to remote servers.
Although Woof is considerably smaller than many cloud-based speech models, it could support applications such as offline voice assistants, personalized podcasts, and device-based accessibility tools. The company has not provided comprehensive independent performance benchmarks comparing Woof with larger commercial systems.
In its company manifesto, Underdog outlines its intention to develop openly available AI models that prioritize local processing and user privacy. The Woof demonstration provides an example of that approach through speech generation running inside a standard web browser.
Featured image credits: X
For more stories like it, click the +Follow button at the top of this page to follow us.