Underdog Squeezes Its Woof Voice Model Into 32 MB and Runs It in the Browser

A voice interface seems like a simple feature until you have to pay for servers, wait for a response from the cloud, and explain to users where their text went. Underdog offers another option: its Woof speech-synthesis model weighs 31.7 MB, contains 31 million parameters, and, according to the company, runs directly in the browser on most devices—even without an internet connection.

The company demonstrated the model not with a benchmark table, but through a small game. The screen shows a room with a typewriter, an old monitor, and a music player. Users type any phrase, and Woof speaks it locally. It is a clear example of what is usually hidden behind the phrase “on-device computing”: text does not need to be sent to a remote server, and the response does not depend on connection quality.

Why Fit a Voice Into 32 MB

With cloud speech synthesis, an application records or accepts text, sends the data to a provider, waits for generation, and receives audio. That chain involves the network, an external API, per-request fees, and data-processing rules. If the service is unavailable or the connection drops, the voice feature stops working.

Woof moves the computation to the user’s device. This approach is particularly appealing for notes, interface narration, navigation, educational applications, and personal assistants. For example, a document-reading app could speak a selected paragraph on a plane or subway without uploading the document’s contents to the cloud. For a product, this means not only privacy but also more predictable costs: once the model has been downloaded, every new request does not become a separate bill from an API provider.

The model is available for download. This gives developers an opportunity to integrate it into their own products, test latency on target devices, and examine its limitations without being tied to a closed service. Underdog previously released Bark, a 167 MB speech-recognition model for local transcription and dictation. Together, these projects outline a complete on-device voice pipeline: Bark converts speech to text, the application processes the request, and Woof speaks the response.

A Demo Is Not Yet a Finished Product

The model’s size is impressive, but by itself it says nothing about how natural the voice sounds, whether Russian is supported, how well it handles long phrases, or how fast it runs on low-end phones. The published material contains no independent comparison of Woof with cloud systems and no detailed measurements of latency, power consumption, or speech quality. The company’s claim that it works on most devices should therefore be treated as a developer statement, not a confirmed test result.

A small model is also unlikely to replace large systems where dozens of languages, precise control over intonation, voice cloning, or studio-quality output are required. Its strength may lie elsewhere: speech that is good enough without a network, a server, or the transfer of personal data.

What Developers Should Test

Start with a small prototype rather than fully replacing the current voice service. Choose one use case—notification narration, for example—and measure first-run time, latency before speech begins, and the load on memory and battery. Then test the pronunciation of numbers, names, abbreviations, and long sentences on the devices the audience actually uses.

If the quality is sufficient, the local model can serve as the default option, with the cloud reserved for complex cases. Woof matters not because the model is smaller than a photo from a modern smartphone. It demonstrates a practical shift: voice AI is gradually becoming an ordinary part of an application—one that does not necessarily need to rent a server or send user data outside the device.