Underdog has released a new demo featuring Woof, a 32MB speech model that operates entirely within a browser without internet connectivity, aiming to democratize voice generation on personal devices.
- Woof: 32MB text-to-speech model runs fully on device in browser
- Model includes 31 million parameters, supports speech and music generation
- Underdog prioritizes privacy with open-source models for edge computing
What happened
Underdog launched a demonstration of its new text-to-speech model called Woof, which is only 31.7MB in size. The demo is a browser-based interactive game featuring a retro-themed room where users can type phrases for the model to vocalize and rotate a tuner to generate music. The model runs locally without requiring any internet connection, emphasizing usability on common edge devices such as phones and laptops.
The Woof model contains approximately 31 million parameters and is publicly available for download. This release follows Underdog’s earlier introduction of Bark, a 167MB speech-to-text model also designed for local device use. The company’s founder, Sigil Wen, highlights a mission to create American-made, privacy-centric AI solutions that empower individuals to retain control over their data.
Why it matters
In an industry largely dominated by large-scale cloud AI models that demand extensive processing power and the continuous transmission of personal data, Underdog’s lightweight models offer a compelling alternative. By enabling sophisticated speech generation and recognition to occur directly on edge devices, users are less dependent on remote servers and can benefit from better privacy, lower latency, and offline functionality.
The ability of a 32MB model to produce both speech and music in real time illustrates that compact, efficient AI systems can provide meaningful experiences even without vast computational resources. This opens doors for wider adoption in personal assistants, content creation, and accessibility tools, especially in areas where internet connectivity is limited or privacy concerns are paramount.
What to watch next
Underdog plans to expand its portfolio of locally run AI models, including a recently announced 27-billion parameter model capable of laptop deployment and an app positioned as a local alternative to popular cloud-based tools like Muse and Instinct. Their future developments may push the boundaries of what edge AI can accomplish in terms of capability and usability.
Industry observers should monitor how Underdog’s open-source approach influences the broader AI ecosystem, particularly in balancing AI performance with privacy and user data sovereignty. Additionally, understanding the trade-offs between these smaller models and traditional large cloud models will be crucial as AI becomes increasingly integrated into everyday devices worldwide.