Bland’s Tofu-1 Brings Full-Duplex Voice AI to Noisy Calls

Bland has released Tofu-1, a full-duplex speech-to-speech model built for phone calls. The San Francisco voice-AI company is offering it as a research preview for a limited number of enterprises, with target sectors including banking, healthcare, insurance and government.

Tofu-1 is designed for calls from moving cars, hospital rooms and weak-signal lines. It listens while it speaks and reasons while it talks, giving phone-agent teams one model for audio in and audio out without a transcript between them.

That design separates Tofu-1 from the usual voice-agent setup. Most voice agents work as cascades: they turn speech into text, pass the text to a language model, then convert the reply back into speech. Tofu-1 listens to audio itself, so tone, hesitation and background sound stay available to the model.

What Tofu-1 hears during a call

Bland lists three audio skills on the product page: speaker tracking, noise separation and paralinguistics. Speaker tracking tells speakers apart and works out who each person is addressing, which matters when several voices enter the same call.

Noise separation separates voices from a crying child, a television, highway traffic, a hospital intercom or keypad tones. Paralinguistics focuses on signals such as hesitation, frustration and urgency, giving the model audio details that a transcript may not capture.

Callers can interrupt or correct Tofu-1 mid-sentence. While it works on a request, it keeps the conversation moving and reports the result only when the work is done. Bland says it never claims a task is finished early.

Because Tofu-1 models the sound of a call, it can also produce one. Its output can include a second speaker, the room around that speaker and line noise, rather than only a clean voice replying in isolation.

Bland has not published the internal architecture. The company describes Tofu-1 as its first full-duplex speech-to-speech model and the first release in a planned family of Bland speech-to-speech models.

The tests, access limits and missing numbers

One demonstration centers on the TV Test, which asks a simple question: can the model follow the caller while ignoring speech from another scene? TV audio is often louder than the caller. Bland marks its TV Test walkthrough as illustrative, not recorded model output.

The company’s only published performance number is self-reported: over 92% could not spot Tofu-1 in screenings. That figure describes the screenings, but it does not answer several questions teams may ask before deployment. Bland has not published size, latency, pricing or benchmarks for Tofu-1.

That leaves a clear gap beside rival full-duplex models, whose published materials include latency and tool-use scores. The comparison figures include 0.798 seconds for FDB v1 turn-taking latency, 448 milliseconds for FDB turn-taking latency and 82.5% tool selection for FDB v3. The comparison list also references OpenAI GPT-Live-1, NVIDIA NemotronLabs and OpenMDW 1.1 for weights.

Tofu-1 access remains a closed enterprise preview, with Bland citing January availability. The sign-up form asks applicants to select monthly call volume ranging from under 1,000 to 50,000+. That detail shows who Bland is trying to reach: teams handling enough calls for interruptions, background noise and audio quality to affect daily operations.

Where Tofu-1 fits in the wider phone-agent market

Tofu-1 arrives as other companies build AI systems around phone calls and live conversations. On November 13, 2026, AI Rudder launched CALL-E, a system that lets AI agents make real phone calls on their users’ behalf. AI Rudder is headquartered in Singapore and has handled more than 80 billion AI calls.

CALL-E works inside ChatGPT, personal AI agents such as Muse, OpenClaw and Hermes Agent, and coding tools including Claude Code, Codex and Cursor. It asks the user before every call and reports back with proof of what was agreed. Its focus is making calls for users, while Tofu-1 focuses on how an AI phone agent hears, reasons and replies during the call itself.

Other announcements point to a broader push toward AI that can participate in live communication. TranGPT is bringing translation, voice synthesis and business integration tools into one platform for cross-border e-commerce, international trade and global business communication. It supports translation across more than 180 languages, along with document translation, real-time voice synthesis and RESTful API integration.

Anthropic has also introduced Claude Meetings, now in beta on Team and Enterprise plans. Claude can join video calls as a visible participant, follow the conversation and start work during the meeting. “Your team can ask Claude questions during the meeting, out loud or in chat,” the company said.

“By default, Claude answers in writing, and it can use your connected tools with your permission.” Claude can also work live on documents, whiteboards and other artifacts while teammates refine them, and “Claude writes the notes and action items, so no one on the call has to.” Support for other video conferencing platforms is coming soon, with Zoom Video, Microsoft and Alphabet among the companies in the video chat space.

For Bland, Tofu-1’s test will come down to the difficult parts of a phone call: competing sounds, interruptions, hesitation and unfinished work. The research preview does not yet include the size, latency, pricing or benchmark details that would allow a full comparison, but it places full-duplex audio at the center of Bland’s next group of speech-to-speech models.

Stay connected via Google News

Follow us for the latest AI News