Underdog demos Woof, a 31.7MB text-to-speech model that runs locally
Underdog has released a browser game built around Woof, a text-to-speech model the company says fits into 31.7MB and runs locally on most devices. With 31 million parameters, it’s small enough for developers to consider shipping speech generation as ...
Underdog’s 32MB speech model puts browser-based voice AI within reach
Underdog has released a browser game built around Woof, a text-to-speech model the company says fits into 31.7MB and runs locally on most devices. With 31 million parameters, it’s small enough for developers to consider shipping speech generation as part of an ordinary web application.
A roughly 32MB download makes local inference easier to justify. Developers still need to weigh that download against a hosted service, and the demo leaves basic questions about performance and compatibility unanswered.
TechCrunch reported the release on November 6, 2026, at 13:02:16 UTC, with an update at 16:57:31 UTC. Underdog’s announcement appeared on November 5.
Trying Woof in the browser
The demo game places users in a retro-styled room with a typewriter, CRT monitor, desk lamp, and music tuner. Type something on the typewriter and the game speaks it back. Rotate the tuner and it generates different kinds of music. Underdog says all the voices use Woof.
Free-form text entry makes this a useful demonstration. Users can try a name, an abbreviation, a long sentence, or an awkward sequence of numbers and hear where speech generation struggles. Curated recordings rarely give them that opportunity.
The game also takes a swipe at “intrusive” cloud providers, reflecting Underdog’s pitch for models that execute on users’ devices and keep their data under local control.
According to the company, Woof can run on most devices without an internet connection, and anyone can download the model. The reporting doesn’t establish compatibility across browser engines, speech quality, or production readiness.
It also describes the tuner generating music without explaining the underlying architecture or separately evaluating that capability. There’s too little detail to infer a general-purpose music-generation system from the game.
Download size and runtime costs
Model download size affects first-use delay, cache storage, bandwidth consumption, and whether someone on a mobile connection will tolerate the feature. At 31.7MB, Woof is small enough to consider loading on demand when a user activates voice output.
That gives product teams a credible option for distributing speech generation as an application asset, reducing the need to maintain a service that handles every utterance.
Runtime memory needs a separate budget. Inference requires working memory, intermediate tensors, audio buffers, and execution machinery. The reporting doesn’t specify Woof’s peak memory use, numerical precision, compression scheme, or browser inference backend.
The combination of 31 million parameters and a 31.7MB package invites questions about weight representation. Without implementation details, developers shouldn’t assume a particular quantization method or treat the downloadable artifact as the entire deployed footprint.
Latency needs testing too. A model can fit comfortably on a device and still take too long to produce audio for an interactive assistant. Developers need at least two measurements: time to first audible output and the rate at which the system produces the remaining speech.
Those measurements should come from target devices. A successful desktop demonstration says little about sustained performance on a lower-end phone, especially when the application is also rendering graphics or doing other work.
Running speech locally
Browser-local synthesis removes the need for a server-side inference request for each piece of text. Computation moves to the user’s device; the publisher still has to distribute the application and model.
For an application that reads short notifications or document passages aloud, that could reduce dependence on a metered speech API. It also shifts the support burden toward heterogeneous processors, browser behavior, battery consumption, and model updates.
An offline reading tool is a useful candidate. Once the necessary assets are downloaded and retained, local synthesis could read sensitive documents without sending their contents to a speech provider. A game could generate dialogue from player-entered text without maintaining a remote synthesis service. Neither use case requires an autonomous assistant or a large language model.
Woof generates speech from text. It doesn’t interpret a request, retrieve information, or decide what to say. An assistant still needs those components, plus speech recognition if users speak their commands.
Underdog has a related component. Weeks before Woof’s demo, it released Bark, a 167MB speech-to-text model intended for edge-device transcription and dictation. Together, Woof and Bark suggest a path toward local voice input and output, though the reporting doesn’t establish an integrated pipeline or end-to-end performance.
Their reported artifact sizes total roughly 199MB before application code, runtime dependencies, caches, or any language model. A complete voice product needs an explicit download and memory budget, even with compact individual components.
Risks and limitations
Evaluation is the biggest gap. The source provides no standardized speech-quality results, latency measurements, device matrix, language-coverage breakdown, or sustained-load testing. Underdog’s claim that Woof runs on “most devices” is too broad to serve as a deployment guarantee.
Teams considering Woof should test their own application text. Product names, technical vocabulary, dates, currency amounts, and long passages often matter more than pleasant delivery of a short sample. A personalized podcast also demands consistency over much longer output than the typewriter interaction demonstrates.
Licensing remains unresolved in the supplied material. TechCrunch says everyone can download the model, and founder Sigil Wen describes a goal of training open models. Neither statement specifies rights for commercial deployment, redistribution, modification, or bundling. Teams need the actual license terms before shipping.
Local speech generation can remove the need to transmit text to a hosted synthesizer, but the surrounding application may still use analytics, remote logging, or other network requests. Teams handling confidential content need to inspect the complete data path and verify offline behavior after initial loading.
Wen’s broader ambition warrants skepticism. He told TechCrunch that “most capable models of today” could run on devices users already own “within months.” A compact speech-generation demo offers no evidence for that prediction.
What developers need before shipping
Underdog’s portfolio extends beyond voice. The company has also announced a 27-billion-parameter model that it says runs on laptops, plus a waitlisted app positioned as a locally hosted alternative to Muse and Instinct. Wen’s manifesto frames the effort around open, American-made models and private AI.
For developers evaluating Woof, the immediate question is whether it can meet a specific product requirement without a cloud dependency.
Start with the license, supported execution paths, offline behavior, and measurements on representative hardware. Then test pronunciation and responsiveness against actual application text.
The 31.7MB package makes Woof worth evaluating. Production use will depend on the engineering details the demo leaves out.
Useful next reads and implementation paths
If this topic connects to a real workflow, these links give you the service path, a proof point, and related articles worth reading next.
Build the data and cloud foundations that AI workloads need to run reliably.
How pipeline modernization cut reporting delays by 63%.
Mistral has released Voxtral TTS, an open source text-to-speech model built for real-time speech generation on edge hardware. The headline specs are strong: nine languages, custom voice adaptation from under five seconds of audio, and latency low eno...
Apple’s Foundation Models framework looked promising at WWDC 2025. Now that iOS 26 is widely deployed, there’s a better read on what it’s actually good for. The first wave of apps tells the story. Developers aren’t stuffing a chatbot into every scree...
Apple's new Foundation Models framework was one of the most important WWDC announcements for developers, even if it didn't get the usual AI stage treatment. The idea is straightforward. Third-party apps can call Apple's built-in foundation models dir...