Overview
Most voice systems on phone lines are still three-stage: turn speech into text, hand the text to a language model, synthesize the reply back into audio. Every hop throws something away. Who was speaking, who a phrase was meant for, whether there was a pause, whether the sentence had finished; all of it collapses into a line of plain text at the transcription step. Tofu-1, which Bland released on Nov 13, tries to skip that middle state and take audio in and put audio out directly. Bland calls it full duplex.
First, what it currently is. A research preview, not something a team can wire into production this month. It is open by application to four kinds of organizations: banking, healthcare, insurance and government. Parameter count, latency, pricing and benchmark scores are all unpublished. So the only thing worth writing about is what it claims to solve, and where to hold reservations.
Bland's own scene makes the problem concrete. Someone calls to report a tree that came down on her house. She is standing in the rain, the signal is bad, her son keeps asking her questions, and halfway through reading out the address she turns away to answer him. Anyone on the other end follows without thinking: she turned aside, wait for her, the rain does not matter, pick up the address where she left off. Current voice systems cannot hold that steady, because transcription erases the turning away and leaves a clean string of words.
Tofu-1 claims to listen to the audio itself, so what the sound carries stays available. Bland splits that into four abilities. It understands the scene, telling speakers apart, working out who a phrase is addressed to, separating a crying child, a television, highway traffic, a hospital intercom and keypad tones from the human voice, and hearing hesitation, frustration and urgency. It generates whole scenes, because it models what a call sounds like rather than only the words, so it can produce a second speaker, the room around them and line noise. It listens while it speaks, so a reply can change mid-sentence when the caller interrupts or corrects it. And it thinks while it talks, running a lookup on identity, a claim or a benefit alongside the conversation.
The most telling of the four is what Bland calls the TV Test. A television is on. An advertisement says call now, a newscaster asks a question, a character says yes. Every phrase is clear and none of it is meant for the agent. Noise removal does not fix this, because the television is not noise. It is speech from the wrong source, and often louder than the caller. Bland's example: the agent asks the caller to confirm she wants to cancel her policy. She says nothing, because she is still looking for the policy number. The television says yes. A transcript-based agent takes that as her answer and cancels the policy. Tofu-1 is built to attribute the yes to the television and ask the caller herself to confirm. Volume cannot settle it. A loud television can be irrelevant and a quiet correction can be the most important thing said on the call.
Bland also states its own limit. If the background makes the caller impossible to understand, the model should say so and ask for a repeat or a quieter place. Ignoring the television must never become an excuse to guess at what the caller said. That line is more candid than most launch material.
The evidence needs separating. Bland says more than 92 percent of listeners in its internal screenings could not tell Tofu-1 output from a real recording, without saying how many listeners, how the clips were chosen, or how the test was run, so treat it as a direction rather than a result. Every voice in the launch video, caller, agent and background, was generated by the model, and Bland labels it as synthetic rather than a real customer call. The TV Test passage is marked an illustrative scenario, not recorded model output. The two samples you can actually hear run 10 seconds and 39 seconds, and both were chosen by Bland.
So the question is not how smooth the demo sounds. It is three other things. Whether it can tell who is speaking on your own line noise, which needs real recordings rather than two clean clips. Whether it admits it cannot hear instead of inventing an answer from context. And whether latency and concurrency hold up on a single call, since none of those three numbers exists yet.
For teams building voice agents, the question Tofu-1 raises is more useful than its answers. The past two years were spent comparing transcription accuracy. This moves the yardstick: on this call, who is speaking, and who is this sentence for. The old cascade architecture had nowhere to put that.
Key Features
- Full-duplex listening and speaking: It hears and speaks at the same time, so a caller can interrupt or correct mid-sentence and the reply in progress changes instead of finishing first.
- Speaker tracking and addressee attribution: It separates speakers and works out who a phrase is meant for. Agent, caller and the child nearby each hold a track, so overlaps are not simply discarded as noise.
- Noise separation: It pulls the human voice out of a crying child, a television, highway traffic, a hospital intercom and keypad tones. Bland stresses that speech from the wrong source is not the same thing as noise.
- Paralinguistic cues: Beyond what was said, it hears how: hesitation, frustration, urgency. These signals largely vanish once audio becomes text.
- Thinking while talking, answers aligned to the latest turn: Verifying identity, checking a claim or confirming a payment runs alongside the conversation. If a caller corrects an appointment date during a lookup, the answer has to reflect the new date, and unfinished work is not reported as done.
- Generating the whole scene: Because it models the sound of a call, it can also produce a second speaker, the space around them and line noise. Every sound in the launch video came from the model.
Use Cases
- Banking identity checks and claim status lookups where a corrected digit must carry through
- Hospital rescheduling when a patient changes the date mid-sentence
- Insurance cancellation or change confirmations with a television or another person in the room
- Inbound and outbound calls from cars, hospital rooms and weak-signal lines
- Voice agent vendors stress-testing whether a model can tell who is speaking on real recordings
Pros
- Processes audio directly, so tone, pauses and who-was-addressed stop collapsing into a single line of text
- Full duplex, so interruptions and corrections land mid-sentence, closer to how two people talk
- Treats the yes from a television as its own kind of problem instead of folding it into noise removal
- Candid about limits, stating openly that it should ask for a repeat rather than guess
- Both published samples are model-generated, so you can listen before deciding rather than reading a description
Pricing
Research preview, with no published price and no self-serve path. The application form asks for monthly call volume in bands from under 1,000 to more than 50,000, limited to banking, healthcare, insurance and government, reviewed case by case. Parameter count, latency, concurrency and benchmark scores are likewise unpublished.
Summary
Tofu-1 is not aimed at transcription accuracy. It is aimed at what transcription throws away. Full duplex lets it be interrupted, speaker tracking lets it work out who a sentence is for, and the TV Test turns a stray yes from a television into a question of attribution rather than noise. It is a research preview open by application to four sectors, with no published size, latency, price or benchmark, and a demo that is entirely synthetic. Whether it is worth waiting for depends on how it handles your own line noise.
Version History
- Research preview released (2026-11-13): Bland released Tofu-1, its first speech-to-speech model, full duplex and processing audio end to end without an intermediate transcript. Four capabilities: understanding the scene (speaker tracking, addressee attribution, noise separation, paralinguistics), generating whole-scene audio, listening while speaking, and thinking while talking. Aimed at banking, healthcare, insurance and government calls, open as a research preview by application, with no published parameters, latency, pricing or benchmark scores. Bland says more than 92 percent of listeners in internal screenings could not tell its output from a real recording, without publishing the method.