23 Daily — According to techcrunch.com, startup Underdog showcased a compact 32MB voice synthesis model on November 6, 2026, capable of generating both speech and music. The model powers a browser-based game called Woof that runs offline on most devices.
32MB Model Integrates Speech and Music Generation
Underdog's Woof features a retro-themed room with a typewriter, CRT monitor, and music tuner. Users can type text which the model converts into spoken words, or adjust the tuner to produce different musical sounds. The speech model has approximately 31 million parameters, is open for download, and runs locally without relying on cloud computing.
Focus on On-Device AI to Protect User Data Privacy
Founder Sigil Wen explained the motivation to build AI models that run on users' devices, aiming to give individuals control over their data and reduce reliance on cloud providers. Underdog previously launched Bark, a 167MB speech-to-text model for transcription designed for edge devices. Additionally, the company released a 27-billion-parameter model operable on laptops and is developing a local alternative app to Muse and Instinct, currently in a waitlist phase.
Editor's take: Small On-Device Voice Models Unlock Diverse Startup Opportunities
Underdog’s lightweight 32MB voice model challenges the norm of large, cloud-dependent AI by enabling private, responsive voice interactions on common devices. For entrepreneurs, this opens possibilities for enhanced personal assistants, customized podcasts, or in-game voice features while minimizing cloud costs and privacy concerns. Startups can validate its value by testing the open-source model’s performance and speech quality across different hardware and assessing suitability for their target applications.