Bland, an artificial intelligence firm based in San Francisco, has released Tofu-1. This is a full-duplex speech-to-speech model currently available in research preview. The system is designed to handle audio from moving vehicles, hospital environments, and connections with poor signal strength. It processes input and output audio simultaneously without generating a text transcript in between.
What is Tofu-1
Tofu-1 listens, reasons, and replies using audio during telephone conversations. It represents the first release in a planned series of speech-to-speech models from the company.
Difference from a cascade voice agent
Most voice agents operate as cascades. They convert speech into text, pass the text to a language model, and then convert the reply back into speech. Tofu-1 listens to audio directly. Tone, hesitation, and background sound remain available to the model. Bland has not published the internal architecture.
Audio understanding
The company lists three audio skills on its product page:
- Speaker tracking: The system distinguishes between speakers and determines who each person is addressing.
- Noise separation: The model isolates voices from crying children, television, highway traffic, hospital intercoms, or keypad tones.
- Paralinguistics: The system detects hesitation, frustration, and urgency.
Full duplex and thinking while talking
Callers can interrupt or correct Tofu-1 mid-sentence. While the system works on a request, it keeps the conversation moving. It reports the result only when the work is complete. It never claims a task is finished early.
Scene generation
Because Tofu-1 models the sound of a call, it can also produce one. That includes a second speaker, the room around them, and line noise.
What is Bland’s TV Test
The TV Test asks one 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.
Comparison with other full-duplex models
| Feature | Bland Tofu-1 | OpenAI GPT-Live-1 | NVIDIA NemotronLabs VoiceChat 11B | Block Dessa |
|---|---|---|---|---|
| Type | Full-duplex speech-to-speech | Full-duplex voice model | Full-duplex speech-to-speech | Full-duplex speech-to-speech |
| Parameters | Not disclosed | Not disclosed | 11B | Not disclosed |
| Weights | Closed | Closed | Open (OpenMDW 1.1) | Not released |
| Training data | Millions of hours of real conversation | Not disclosed | ~550k hours audio plus text | Human speech only, no text |
| Hardware | Bland-hosted | OpenAI-hosted | A100, H100, H200, B100, B200, RTX-6000 | MacBook Pro M4 Max or better |
| Turn-taking latency | Not disclosed | 0.798 s (FDB v1) | 448 ms (FDB) | Not disclosed |
| Tool use | Keeps talking while it works | Delegates to backend model | Live tool calling, 82.5% tool selection (FDB v3) | Not disclosed |
| Access | Enterprise research preview | API and ChatGPT | Hugging Face checkpoint, NVIDIA container | Research preview for safety researchers |
Competitor figures come from each vendor’s own page. FDB results use different benchmark versions and are not directly comparable.
How can teams access Tofu-1
Research preview access is open to a limited number of enterprises. Bland names banking, healthcare, insurance and government as target sectors. The sign-up form asks for monthly call volume, from under 1,000 to 50,000+.
What it means for users
Teams using phone agents will find a single model handling both input and output audio. This removes the need for a transcript between the two stages. For people in noisy environments, the system can separate the voice from background sounds like traffic or intercoms. Users can interrupt the system mid-sentence without breaking the flow of conversation.