AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Underdog shows off a 32MB speech model in new demo

Collected Nov 6, 2026

Underdog, a startup founded by Sigil Wen, has released a small text-to-speech model called Woof and packaged it inside a browser game of the same name, the company announced on November 5, 2026. The model is 31.7 MB in size and has 31 million parameters, according to the company, and It is available for anyone to download.

The game is set in a room with a retro tech theme: a typewriter, a CRT monitor and a desk lamp whose switch makes a sound when clicked. A music tuner can be rotated to generate different kinds of music. Typing on the typewriter makes the game speak the text back through the Woof model. The intro warns players about "intrusive" cloud providers in the town, an apparent jab at models that rely on remote servers. All of the voices in the game are produced by Woof, and the company says finishing it earns an invite.

The central claim is that the model runs on most devices through the browser with no internet connection at all. That is the interesting part for developers: a text-to-speech model small enough to ship as part of a web page changes the deployment math. Today, speech generation usually means calling a hosted API, paying per character or per minute, and accepting the round-trip latency and privacy cost of sending text to someone else's server. A model measured in tens of megabytes can be cached by the browser, executed locally and used offline, which removes the per-request bill and keeps the input on the device. Whether the audio quality holds up at this scale is a separate question that a demo game answers only partly.

The Woof release follows Bark, a 167 MB speech-to-text model that Underdog launched weeks earlier, also aimed at edge devices for transcription and dictation. Together the two point at a local voice stack: speech in and speech out, both running on hardware the user already owns. Underdog has also released a 27-billion-parameter model that it says can run on laptops, plus a waitlisted app positioned as a locally hosted alternative to Muse and Instinct.

Wen told TechCrunch he began building because he wanted his own system rather than using notetakers such as Granola, with models that run on users' devices and let them keep their data. He said he expects most of today's capable models to run on devices people already own within months, and described an aim to train open, American-made models that allow private AI, referring to his manifesto.

A 31-million-parameter model is roughly three orders of magnitude smaller than the 27-billion-parameter model the same company says runs on laptops, so the quality ceiling is clearly different from frontier systems. Small speech models typically trade expressiveness, prosody and voice variety for size, and the demo's range, from spoken typed text to generated music, suggests broad coverage rather than studio-grade output. That is a reasonable guess from the parameter count, not a published benchmark; the company has not described evaluation results here. In practice, a model this size is a fit for short utterances, UI feedback, offline assistants and personalized audio rather than for replacing production voice actors or large hosted TTS services. Generating music from a 31M-parameter speech model is a striking claim, and the honest reading is that it likely produces short, stylized fragments rather than full compositions.

It is also worth separating the model from the marketing artifact. The game is a demonstration designed to prove a point about running locally, and the invite tied to completion is a distribution mechanic. The downloadable weights are the durable part.

Why it matters: Teams building voice features now have a concrete example of a downloadable, offline-capable text-to-speech model small enough for a browser, which makes local voice interfaces and privacy-preserving audio plausible without a cloud bill. This suggests the near-term impact is on edge assistants, accessibility features and personalized audio, not on replacing large hosted models, and developers should expect to trade some audio quality for the savings in cost, latency and data exposure. The 27-billion-parameter laptop model and the waitlisted app are worth watching as signs of how far the local-first approach can scale.

Read at TechCrunch · AI

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Despite its small size, Underdog's new model can generate both speech and music.