
Voice-AI company Bland has opened an enterprise research preview for Tofu-1, a full-duplex speech-to-speech model designed for live telephone workflows in sectors such as healthcare, insurance, and banking. Unlike conventional voice pipelines that chain speech-to-text, an LLM, and text-to-speech, Tofu-1 processes and generates audio directly without an intermediate text transcript.
Full-duplex handling and noisy environments
According to Bland's announcement, removing the text bottleneck lets the model listen and speak simultaneously, so callers can interrupt or redirect the agent mid-sentence while background tool calls or database lookups run in parallel. The architecture is also tuned to preserve acoustic cues such as hesitation and urgency and to separate speaker audio from car, home, or hospital background noise.
Availability and verification limits
Tofu-1 is currently restricted to an application-based closed enterprise research preview on Bland's platform. No public API pricing, open weights, latency benchmarks, or independent evaluations in regulated environments have been published. For clinical or safety-sensitive phone lines, real-world reliability and compliance remain unverified outside vendor testing.
Original reporting; sources reviewed November 16, 2026. Closed research preview; no hands-on testing claimed.



