Startup Underdog has demonstrated a text-to-speech model small enough to run locally inside a web browser. Woof is a 31-million-parameter model packaged in 31.7MB, and the company has released its weights for anyone to download.
A browser game lets users type a phrase and hear the model speak it without sending the request to a cloud service. A second control generates different kinds of music. Underdog says the model can operate on most devices without an internet connection, making it a possible building block for private assistants, games or personalized audio.
The company is presenting Woof as an edge model, not a replacement for much larger cloud systems. TechCrunch’s report does not include an independent comparison of voice quality, speed or language coverage, so the practical limits remain unclear.
Woof follows Bark, Underdog’s 167MB speech-to-text model for local transcription and dictation. Together they point to a different deployment trade-off: accepting a narrower model in exchange for offline use, lower infrastructure demand and keeping audio data on the user’s own device.