Bland, the San Francisco voice-AI company, has released Tofu-1, a full-duplex speech-to-speech model built for phone calls where the audio is already bad. Announced on 13 November and widely reported on 15 November, it is the first release in a planned family of Bland speech-to-speech models, and it is open now as a gated research preview for enterprises.
Unlike most voice agents, Tofu-1 processes audio directly instead of transcribing it into text first. It listens and speaks at the same time, so a caller can interrupt or correct it mid-sentence, and it is designed to separate the caller's voice from whatever is happening in the background — a television, a crying child, traffic, a hospital intercom or keypad tones. Bland says it can also tell speakers apart and work out who a phrase is addressed to, so a TV saying “yes” in the background does not accidentally cancel an insurance policy.
The preview is deliberately narrow: access is open to a limited number of enterprises in banking, healthcare, insurance and government, with general availability slated for January. Bland has published no benchmarks, model size, latency figures or pricing, and its most eye-catching claim — that over 92% of people in screenings could not spot the model — is self-reported, not independently verified.
Tofu-1 lands two days after inclusionAI's Duo v1, another native full-duplex launch, as voice-AI companies race toward agents that can hold a real, interruptible phone conversation. Until independent tests arrive, Tofu-1 is best read as a direction of travel: the phone call, not the text box, is becoming the next frontier for AI agents.