As artificial intelligence companies continue to compete over increasingly powerful models, a startup is taking a different approach by shrinking AI technology enough to run directly on everyday devices without relying on cloud servers.
Underdog, an AI startup, has unveiled Woof, a lightweight text-to-speech model measuring just 31.7MB, demonstrating how smaller AI systems could make voice generation more accessible, private and less dependent on internet connectivity.
The company introduced the model through an interactive browser game designed to showcase its capabilities. The demonstration features a retro-themed room fitted with a typewriter, a CRT monitor, a desk lamp and a music tuner.
Players can interact with different objects in the room, including switching on the lamp to trigger a sound. The typewriter allows users to enter any text using their keyboards, which Woof then converts into spoken audio. A music tuner also lets users generate different musical sounds.
The game includes a fictional storyline about defeating a character called Darkcloud, an apparent criticism of dependence on centralised cloud computing services.
According to the company’s November 5 announcement on X, every voice featured in the game runs on Woof, with the entire experience operating inside the browser.
Underdog says the model can function on most devices without an internet connection. It contains approximately 31 million parameters and is available as an open model for developers and other users to download.
Small AI models could change how people use voice technology
Although Woof is unlikely to compete directly with larger speech-generation systems in terms of overall capabilities, its relatively small size highlights an important development in AI: the growing push to move artificial intelligence processing away from remote servers and onto users’ personal devices.
This approach, commonly known as edge AI, could reduce dependence on cloud infrastructure, improve response times and give users greater control over their information.
For markets such as Nigeria, where internet reliability, data costs and access to expensive computing infrastructure remain considerations for technology adoption, lightweight AI models could open new possibilities.
Developers could potentially build offline voice assistants, transcription applications, educational tools and personalised audio experiences without requiring users to maintain constant internet connections.
However, the practical benefits will depend on how well these smaller models perform across different devices, languages and accents. For Nigerian developers, support for local languages and Nigerian English will also be important before such technology can achieve wider adoption.
Underdog expands its portfolio of lightweight AI models
Woof is not the startup’s first attempt to make AI models smaller and more accessible.
Weeks before unveiling the text-to-speech model, Underdog released Bark, a 167MB speech-to-text model designed to operate on edge devices. The model targets transcription and dictation, allowing users to convert spoken words into text without depending entirely on cloud-based processing.
The startup is also developing a 27-billion-parameter model capable of running on laptops, alongside a waitlisted application positioned as a locally hosted alternative to existing AI productivity tools.
Underdog founder Sigil Wen has previously explained that his interest in on-device AI grew from wanting to build a personal system that would give users greater ownership of their data rather than relying on third-party note-taking services such as Granola.
In an interview previously reported by a technology publication, Wen said he believed increasingly capable AI models would soon run on devices people already own.
He also emphasised his ambition to develop open, capable models that support private AI experiences.
The company’s approach reflects a growing argument within the AI industry: the future of artificial intelligence may not depend exclusively on building larger models, but also on making existing capabilities smaller, cheaper and easier to run.
While cloud-based AI services will continue to play an important role, projects such as Woof suggest that everyday computers and smartphones could eventually handle more sophisticated AI tasks independently.
For Underdog, the challenge now will be proving that its compact models can deliver consistent performance beyond impressive demonstrations. For developers, however, the prospect of running voice-generation technology in a browser with less than 32MB of model data is already worth watching.
