Introduction

Voice-agent company Bland has unveiled Tofu-1, its first speech-to-speech model. Bland

on November 14 (about 02:03 MYT), and the Tofu-1 research page went live at the same time. Tofu-1 is a full-duplex audio model: it hears and speaks at the same time, with no text step in between. Bland built it for phone calls made in bad conditions, like from a moving car, a hospital room, a weak signal or a house with the TV on. It is closed. For now the only way in is an enterprise research preview you request through a form on that page. There are no weights, no public API and no pricing yet.

This is a preview piece, not a "go try it" piece. Still, it's worth a close look. The launch video from CEO Isaiah Granet passed 1.4 million views in under two days. Bland also says every sound in its launch video was generated by Tofu-1: the caller, the agent and the background noise.

What Bland Announced

  • Model: Tofu-1. Bland calls it "the first release in Bland's family of speech-to-speech models."
  • Modality: audio in, audio out, full duplex. There is no speech-to-text, LLM, text-to-speech cascade in the middle.
  • Understands audio: it tells speakers apart, works out who each phrase is meant for, separates voices from noise, and picks up hesitation, frustration and urgency.
  • Generates whole scenes: it can produce several speakers, overlapping voices and the room around them. One page sample is a 10-second clip of a man calling from his car while his passenger talks to him, all generated.
  • Listens while it speaks: callers can interrupt or correct it mid-sentence. It is meant to treat a "mm-hmm" as keep going and "No, it's 4, not 9" as stop.
  • Thinks while it talks: the conversation and the background work (identity checks, claim lookups, payments) run side by side. A 39-second sample shows an agent that keeps talking during a lookup and gives the result only once it's done. Bland stresses that the model "never says a task is done before it is."
  • Training data: Bland says it was trained "on the callers that enterprises talk to every day."
  • Bandwidth: the spectrograms on the page top out at 4 kHz, which is phone-line audio, not studio-grade output.

The "TV Test"

The most useful idea on the page is what Bland calls the TV Test:

  • The agent asks: "Just to confirm, you want to cancel your policy?"
  • The caller says nothing. She is still looking for her policy number.
  • The TV in the room says: "Yes!"

A cascaded agent hears a clean "yes" in the transcript and might cancel the policy. Tofu-1 is built to attribute that "yes" to the television and ask the caller to confirm. Noise suppression can't fix this case, because the TV isn't noise. It's clear speech from the wrong source, and it's often louder than the caller. Bland also sets a limit: if the background makes the caller impossible to understand, the model should ask them to repeat themselves or move somewhere quieter, not guess.

One caveat: Bland labels this scenario as illustrative, not a recorded model output. The interruption and "two tracks" timing diagrams are also marked as conceptual. The only real outputs on the page are the two audio clips.

Where to Run It Today

  • Research preview only: request access on the Tofu-1 page. The form asks for your work email, phone, website and monthly call volume (from under 1,000 to 50,000+). Bland says access is open to "a limited number of enterprises" and that it will email people as slots open.
  • Who Bland wants: teams that handle high-stakes calls in banking, healthcare, insurance and government, and that want to find where the model breaks.
  • What you can do right now: listen to the two samples on the page and watch the
    . There's no playground, SDK or API endpoint yet.

Caveats

  • No numbers. There are no latency, WER, turn-taking or interruption benchmarks, no model size and no paper. Every capability claim comes from Bland's own page and videos.
  • Demo audio is vendor-picked. The fully generated launch video shows what the model can make. It doesn't show how well it understands a real noisy call.
  • Closed and gated. No weights, no pricing and no self-serve access. Indie makers and small studios can't use it yet.
  • Phone-band audio. At 4 kHz it's tuned for telephony. Don't expect it to stand in for a studio TTS voice in a game or film.
  • Crowded lane. Full-duplex voice has been busy this month. If you want something you can run today, look at the open-weight Haru-4B, Sophon Realtime v1 and MOSS-Live, or at hosted speech-to-speech options like Phonic-1 and Rime Prism. What sets Tofu-1 apart is its focus on who is speaking, and to whom, on a noisy line.

Conclusion

Tofu-1 puts a name on a problem every voice-agent builder runs into: the TV saying "yes," the passenger cutting in, the caller who turns away mid-sentence. A cascade that flattens all of that into a transcript gets these moments wrong. A model that keeps the sound itself has a chance to get them right. The scene-generation side is quietly interesting too. A model that can generate a caller, an agent and the rain around them could produce realistic noisy test calls for evaluating other agents. Until Bland shares benchmarks or opens access more widely, treat Tofu-1 as a strong design direction, not a tool you can ship with. If your team runs real phone lines in banking, health or insurance, the preview form is the place to start.

—Anabel ♡