AI startup Underdog has released a new text-to-speech model called Woof. The model is 31.7MB in size and is designed to generate speech directly on a phone or computer without relying on cloud servers.
According to Underdog, Woof contains approximately 31 million parameters. The company states that the model can run on most devices without an internet connection once the necessary files are available locally. This allows speech generation to occur on a user’s device rather than requiring text to be sent to a remote server. Underdog has made the model available for anyone to download.
The company notes that while Woof is lightweight, its small size does not necessarily mean it can match the capabilities or quality of much larger AI systems. However, its size could make it suitable for applications such as text-to-speech tools, offline voice assistants, and personalized audio content.
To demonstrate the technology, Underdog created a free, interactive browser game called Escape on November 5, 2026. The game features a retro-themed room where players can type messages using a virtual typewriter, which Woof then converts into spoken audio. The game also includes objects like a music tuner that produces different sounds. The introduction of the game refers to intrusive cloud providers, reflecting the company's focus on local AI.
Woof follows Underdog’s previous release of Bark, a 167MB speech-to-text model for transcription and dictation on local devices. The startup has also introduced a 27-billion-parameter AI model for laptops and is developing an application to allow users to interact with locally hosted AI models as an alternative to cloud-based assistants.
Underdog founder Sigil Wen told TechCrunch that he started developing these tools because he wanted an AI system that could run on his own devices and give users greater control over their information. The company's goal is to reduce dependence on cloud computing while making capable AI tools available on existing consumer hardware. While larger models still offer advantages for complex tasks, Underdog's release shows how smaller models can handle specific functions like speech generation without remote infrastructure.
