Underdog's Woof is a 32MB speech synthesis model that runs fully offline in the browser

Underdog has demonstrated Woof, a speech synthesis model that occupies just 32MB and carries 31 million parameters, according to TechCrunch. The model runs entirely on the user's machine inside the browser — no part of the interaction touches the network — and its weights are published for open download, so developers can pull the model down and run it inside their own applications and pages.

From cloud service to shipped asset

That footprint is unusual for voice. Speech features on the web have generally been built on hosted APIs: text is sent to a remote server, synthesis happens there, and the resulting audio travels back. That arrangement adds latency to every request, bills on each call, and routes user content through third-party infrastructure.

A model this small inverts it. Audio is produced on the device in front of the user, there is no per-call inference cost, and the feature keeps working when the connection drops. In a browser context, synthesis stops behaving like a network dependency and starts behaving like any other static resource — something that ships with the frontend and can be cached and versioned alongside the rest of the app's files. Because the weights are open, developers can embed the capability without wiring up a vendor relationship or a metered API key.

Part of a pattern at Underdog

Woof is not an isolated release. Weeks earlier, per TechCrunch, the company shipped Bark, a 167MB speech-to-text model intended for transcription at the edge. Underdog has also released a 27-billion-parameter model designed to run on laptops.

Taken together, the three releases sketch a consistent thesis: capable models can sit on the user's own hardware rather than in a datacenter, and a single lab can address speech recognition, speech output and larger-scale reasoning across both browsers and laptops. The common ingredient is packaging — open weights at sizes an individual developer can download and get running locally the same day. There is also a distribution effect: models compact enough to travel inside an application can reach users without dedicated inference infrastructure.

What the reporting leaves open

The coverage centres on size, licensing and distribution rather than benchmarks. How Woof's output compares with large hosted voice systems in naturalness and robustness is not addressed in the reports, and that comparison will decide how far client-side synthesis can be pushed beyond demos and simple interfaces. A 31-million-parameter model is tiny by current standards, and the trade-offs involved in that compression are not detailed.

Why it matters

For builders, a downloadable 32MB voice model turns speech from a backend integration into a frontend feature. There is no inference bill, latency drops because nothing crosses the network, and audio stays on the device by default — which quietly simplifies the privacy conversation, since there is no user audio to disclose or secure in transit. Paired with Bark for on-device transcription, a small team could assemble a complete offline voice loop built on open weights.

More broadly, Woof is another data point in the migration of AI capability from the datacenter to the client. Speech synthesis was, until recently, a textbook example of a task that required a server. If it now fits in a browser download, the list of features that genuinely need cloud inference keeps shrinking — and voice interfaces, long constrained by per-call costs and latency budgets, become something a single HTML page can carry.