A 32MB Voice Model Tests Whether Speech Can Leave the Cloud

Quick take

Underdog AI’s Woof has 31 million parameters and a footprint of roughly 32MB.
Its one-voice version runs offline in a browser, generating speech without sending text to a remote API.
Underdog’s model card lists a 40.6MB two-voice version alongside the 31.7MB one-voice browser model, under a non-commercial license.
Luis Chavez-Mattos found Woof handled simple sentences but stumbled over dense numbers, names, fractions and long passages.

Why it matters

The practical change is that developers can test voice features without a remote API, but they inherit responsibility for browser integration, pronunciation failures and licensing limits. The next question is whether Woof can handle the numbers, names and long passages that private assistants, accessibility tools and other real-world uses require.

Underdog AI’s Woof makes local text to speech small enough for a browser, but its real test is not whether it can speak. It is whether people can trust what it says.

For a person using a phone in a poor connection area, or a company handling sensitive notes, cloud voice services create an awkward tradeoff. They are convenient and increasingly natural, but text must leave the device before it comes back as speech. Woof, a new model from Underdog AI, is an attempt to remove that compromise.

Model footprints in MB; the model has 31 million parameters. — see the chart in the original article.

TechCrunch reported that the company is showcasing a text to speech model with 31 million parameters and a footprint of roughly 32MB. That is not a frontier system designed to compete with the most expressive commercial voices. Its significance is more practical. A model this small can fit on a phone, laptop or embedded device, where offline operation, privacy and response time may matter more than theatrical realism.

The company’s official Woof demo presents the one-voice version as an offline, in-browser experience. The browser does not need to send text to a remote API for generation. That changes the economics and architecture of a voice feature: there is no per-request cloud bill, no dependency on a live connection and less exposure of what a user types.

The release is slightly larger and more complicated than the headline number suggests. Underdog’s model card documents a two-voice version measuring 40.6MB, alongside the 31.7MB one-voice browser model. It also lists the model’s parameter counts, CPU performance, memory use, benchmark results, usage instructions and audio samples. The model is available under a non-commercial license, an important limitation for developers considering it for products rather than experiments.

That distinction matters because small models are valuable only when they are usable. A developer needs more than downloadable weights. They need predictable output, acceptable speed, clear licensing and a runtime that does not require rebuilding half the speech stack.

TTSLab, which independently ported Woof to TypeScript and WebGPU, praised the reference implementation as readable and its output as deterministic. That work suggests the model can be adapted to modern browser hardware rather than remaining a research artifact. But TTSLab also noted that Woof ships weights and reference code, not a polished production runtime. For an engineer, that can turn an impressive demo into a project involving audio pipelines, browser compatibility and performance tuning.

The sharper criticism comes from hands-on testing. Luis Chavez-Mattos reported in his review that Woof handled simple sentences well, but stumbled over dense numbers, names, fractions and long passages. Those are not obscure edge cases. They appear in news articles, financial documents, calendars, addresses and accessibility tools, precisely the kinds of places where speech generation is expected to be dependable.

Chavez-Mattos also warned that the benchmark is self-reported and that the non-commercial license could prevent commercial deployment. That is the central challenge to the small-model story. A voice that is private and nearly instant is not necessarily useful if it mispronounces a customer’s name, reads a price incorrectly or requires a developer to work around restrictive terms.

Woof therefore sits in an uncomfortable but productive middle ground. It is more concrete than a claim that edge AI is coming, because people can run the model locally in a browser. Yet it is not a replacement for cloud systems that offer broader voice selection, stronger pronunciation handling and production support.

The comparison with Underdog’s larger Bark speech-to-text model makes the tradeoff clearer. Larger systems can pursue richer capabilities, while Woof narrows the task to a compact, constrained experience. That may be exactly the right design for private assistants, educational tools, local notifications or devices that cannot rely on a server.

The broader question is not whether a 32MB model can make sound. Woof has already demonstrated that. The question is whether developers and users will accept modest expressiveness and occasional errors in exchange for control. If they do, tiny speech models could make offline voice a normal product feature rather than a technical curiosity. If they do not, Woof will remain an elegant demonstration of what is possible, but not yet a reason to leave the cloud.