A Deterministic, Deficit-Aware Framework for Egocentric Data Collection
Published August 15, 2026
Read article →01 / Mission
Building voice infrastructure.
Concentrated fully duplex models.
Bellu.ai is a voice lab. We train concentrated fully duplex models — systems that listen and speak at the same time.
02 / Duplex
Concentrated fully duplex models
Most voice AI is half-duplex: wait, then talk. Humans overlap, interrupt, and breathe. We train concentrated fully duplex models — focused, not giant chat models with TTS glued on — that listen and speak on the same stream.
Half-duplex
ListenWait
SpeakWait
Turn-taking. Barge-in is a hack. Latency stacks across STT, LLM, TTS.
Full duplex
Listen
Speak
Simultaneous listen and speak. Interrupts are native. One concentrated model, one clock.
/ Concentrated
Narrow models for live conversation — not general assistants wrapped in audio.
/ Barge-in
User speech is always in. The model yields without a separate VAD pipeline.
/ Same stream
Listen and speak on one audio path. No round-trip glue.
03 / Research
Lab notes
Open research folder →The Architecture of Embodiment: Multimodal Data Standardization and Management for Generalizable Embodied AI
Published November 14, 2025
https://www.techrxiv.org/users/998761/articles/1360379-the-architecture-of-embodiment-multimodal-data-standardization-and-management-for-generalizable-embodied-ai04 / Contact
Partnerships on voice infrastructure and duplex models. Direct to the founder.
santhosh@belluai.comSubject: VOICE_INFRASTRUCTURE