AI firms globally are in competition to develop larger and more advanced models, but they face the challenge of requiring substantial processing power. Conversely, some innovators and startups are focused on creating smaller models that can operate on devices such as smartphones and computers.
A notable startup, Underdog, has launched a game utilizing its compact voice model named Woof, which is only 32MB in size.
Set in a room themed with retro technology, the game includes items like a typewriter and a CRT monitor, alongside a desk lamp that makes a sound when clicked. The introductory part of the game cautions users about “intrusive” cloud services existing in the locality, subtly critiquing cloud-based solutions.
Players can engage with both the typewriter and a music tuner.
Typing on the keyboard will prompt the game to vocalize the text, while turning the tuner allows players to create various music styles.
This interaction takes place directly in the browser using Underdog’s speech generation model, which the company claims can function on most devices without needing an internet connection. With 31 million parameters, this model is freely available for download.
While it’s unlikely that such a compact model will fully replace larger ones, it still has the potential to enhance local assistant capabilities or enable personalized podcast creation.
This demonstration follows Underdog’s earlier introduction of Bark, a 167 MB speech-to-text model also designed for edge devices, aimed at transcription and dictation tasks.
Sigil Wen, Underdog’s founder, shared with TechCrunch that his motivation for creating the app stemmed from a desire to establish an AI system that allows users to control their own data, rather than relying on services like Granola for note-taking.
“I believe that within months, we will be able to run many of today’s advanced models on devices that people already own. Our goal is to develop the best open, American-made models that prioritize user privacy,” he explained, describing his vision.
Alongside the two voice models mentioned, Underdog has also introduced a 27-billion-parameter model that can function on laptops, as well as a waitlisted application intended to serve as a locally-hosted alternative to existing platforms like Muse and Instinct.
