Underdog shows off a 32MB speech model in new demo

Most of the industry is racing toward larger models and bigger data-center bills. A parallel track is compressing useful capability onto phones and laptops so speech never has to leave the device.

Startup Underdog released a browser game built on Woof, a roughly 31.7MB text-to-speech model with about 31 million parameters. Type on a virtual typewriter and it speaks back; twist a tuner for music. The company says it runs on most devices without internet, and the weights are downloadable. It follows Bark, a 167MB speech-to-text model for on-device transcription.

If local voice gets good enough for assistants and personal media, privacy stops being a slogan and becomes the default architecture.

We're rooting for this lane. Not every voice job needs a frontier datacenter model. Reading a note aloud, dictating a memo, or generating a private podcast draft are exactly the jobs where a 32MB model that never phones home is the right product.

The demo's aesthetic is the message: retro desk, CRT, a jab at "intrusive" cloud providers, defeat Darkcloud for an invite. Founder Sigil Wen told TechCrunch he started building because he wanted his own notetaking stack instead of tools like Granola — American-made open models that run on devices people already own, with data that stays theirs. Underdog has also released a 27-billion-parameter model aimed at laptops and a waitlisted local app positioned against Muse and Instinct.

I'm clear-eyed about limits. A 31-million-parameter TTS model will not match the expressiveness or robustness of large cloud voices across languages, noisy conditions, and long-form narration. TechCrunch is right that it will not replace bigger models wholesale. The useful claim is narrower: good enough for local assistants and personalized audio, with a path to improve without surrendering the mic stream.

If this works and spreads, the upside is real agency — offline accessibility tools, on-device dictation in regulated workplaces, kids' toys that do not become always-on surveillance. The downside if quality stalls: local AI becomes a niche cosplay while cloud defaults keep winning on convenience. We're watching whether Woof and Bark become developer defaults or stay clever demos.

Edge AI has accelerated as phones and PCs gain NPUs, and as privacy rules and enterprise data-residency demands make cloud-only voice harder to justify. Underdog is pitching into that gap with open weights and a consumer waitlist.