Skip to main content

Out of Set

Category: Voice / Speech AI

South Korean on-device speech AI startup building edge-native foundation models for voice recognition and synthesis, designed from the ground up for device-level deployment without cloud dependency. Out of Set was founded in 2026. The company is led by Hyeongju Kim (김형주). Based in Seoul, South Korea. Team size: 1-10. Latest round: Seed. Key investors include TheVentures (더벤처스).

Founded
2026
Headquarters
Seoul, South Korea
Team size
1-10

Value proposition

Edge-native on-device speech AI models that are task-perfect, zero-overhead, and instantly deployable — designed from scratch for device environments rather than compressed from cloud models, enabling privacy, low latency, and offline reliability without performance degradation.

Products and solutions

On-device speech AI models (voice recognition/STT and voice synthesis/TTS); Voice Agent (hands-free conversational AI); Live Translation (real-time speech-to-speech); Instant Captions (live subtitles, meeting notes, dictation); Voice Transformation (enhance quality, change speaker, adjust accent)

Unique value

Unlike competitors who compress large cloud models (taking months to deploy with quality loss), Out of Set builds native on-device foundation models from the ground up — delivering best-in-class performance that deploys instantly to PC, mobile, wearable, or IoT without battery drain, frame drops, or memory pressure.

Target customer

Security-sensitive verticals (healthcare, legal, financial); smartphone, automotive, and robotics OEMs; enterprises needing private, low-latency, offline voice AI

Industries served

Healthcare; Legal; Financial Services; Smartphone; Automotive; Robotics

Technology advantage

Native on-device foundation model architecture designed from scratch for edge deployment; proprietary automatic fine-tuning system enabling rapid customer-specific model adaptation; ultra-lightweight vertical AI models that run without internet connection; proven technical excellence via #1 HuggingFace ranking and thousands of GitHub stars on prior open-source work

How they differentiate

Native edge-first architecture (not compressed cloud models); automatic fine-tuning system for customer-specific adaptation with fewer resources; founded by top-tier speech AI researchers with proven open-source track record (HuggingFace #1, thousands of GitHub stars); zero-overhead design that doesn't compete with host app resources

Main competitors

Selvas AI (셀바스AI) — Korean on-device STT/TTS provider; Supertone (HYBE) — on-device TTS (Supertonic); Nuemath (뉴메스) — lightweight foundation models

Major milestones

Founded April 2026; Seed investment from TheVentures (May 2026); CEO's prior open-source on-device TTS model ranked #1 on HuggingFace with thousands of GitHub stars

Market positioning

Early-stage Korean startup in the on-device speech AI space, competing against established Korean players (Selvas AI) and HYBE-backed Supertone. Differentiates via native edge-first architecture vs. compressed cloud models. Targeting security-sensitive verticals first.

Geographic focus

South Korea (primary); global markets via open-source community

About Hyeongju Kim (김형주)

Ex-Neosapience Speech AI Developer; Ex-Supertone (HYBE) Speech AI Lead (5 years total speech AI development). M.S. in Electrical & Information Engineering, Seoul National University. Published open-source on-device TTS model that ranked #1 on Hugging Face in its category with thousands of GitHub stars.

Official website: