Underdog demos a 32MB speech model that also makes music

The AI industry is still largely focused on building larger models, but a parallel effort is pushing in the opposite direction: making models small enough to run directly on phones, laptops, and other edge devices. Underdog, a startup focused on open local AI, has used a browser game to showcase its new speech-generation model, Woof.

Introduction: a speech model under 32MB

Woof is approximately 31.7MB in size and has around 31 million parameters. The demo takes place in a retro-styled room containing a typewriter, a CRT monitor, a desk lamp, and a music tuner. Users can type text with a keyboard and have the virtual typewriter read it aloud. Turning the tuner produces different kinds of music. The voices and audio in the experience are generated by Woof.

Underdog says the model runs in the browser and can work on most devices without an internet connection. It has also made the model available for download. If that claim holds across a broad range of hardware, developers could build basic voice interactions without sending every request to a cloud service.

Key points

• Very small footprint: At about 31.7MB, Woof targets a different operating model from server-heavy speech systems.
• More than text-to-speech: The demo includes speech and music generation, although the source does not provide detailed benchmarks for quality, latency, or language coverage.
• Local-first design: Running at the edge could reduce network dependence and give users greater control over voice data.
• Part of a broader effort: Several weeks earlier, Underdog released Bark, a 167MB speech-to-text model intended for transcription and dictation on edge devices.

Why it matters—and what remains unclear

The strongest case for local speech AI is privacy. Meeting notes, personal reminders, and accessibility features could potentially be processed on the device rather than continuously uploaded to a remote service. Offline operation may also be useful where connectivity is unreliable or cloud inference is too expensive. For developers, a compact model is easier to package inside a browser experience, desktop application, or consumer device.

Size alone, however, does not establish practical superiority. Independent testing is still needed to determine how Woof handles long passages, difficult names, multiple speakers, different languages, and sustained generation. Underdog has not presented it as a replacement for larger models. A more measured interpretation is that Woof could become a building block for local assistants, personalized podcasts, and lightweight entertainment products.

Woof and Bark show Underdog pursuing a consistent strategy: combine open models with local inference and let users retain more control over their data. The company has also discussed a larger model that can run on laptops and a locally hosted application still on a waitlist. As device hardware improves, the key challenge will be balancing audio quality, response speed, and privacy well enough to move compact speech models from demos into everyday software.

Source: TechCrunch AI