Most voice demos are recorded in quiet rooms. Real phone calls happen in cars, hospital hallways and kitchens with a TV running. Bland is betting that this gap is where voice agents fail, and its new model, Tofu-1, is built around it.
The San Francisco company released Tofu-1 as a research preview. It is a full-duplex speech-to-speech model, so it listens while it speaks and keeps reasoning mid-sentence. Callers can interrupt or correct it. One model takes audio in and puts audio out, with no transcript in between.
That last part is the real shift. Most production voice agents are cascades: speech-to-text, then a language model, then text-to-speech. Each hop adds latency and strips information. Tone, hesitation and background sound vanish at the transcription step. Tofu-1 hears the raw audio, so those signals stay available. Bland has not published the architecture, though, so how it does this is unknown.
Bland lists three audio skills. Speaker tracking tells voices apart and works out who is addressing whom. Noise separation pulls a voice out of a crying child, highway traffic, a hospital intercom or keypad tones. Paralinguistics picks up frustration, urgency and hesitation. Bland also says the model keeps the conversation moving while it works on a request, and only reports a result once the work is actually done. Because it models the sound of a call, it can also generate one, including a second speaker, the room and line noise.
Bland frames its benchmark as the TV Test: can the model follow the caller while ignoring speech from another scene? TV audio is often louder than the caller, so it’s a fair stress case. But Bland marks its walkthrough as illustrative, not recorded model output. That’s not evidence.
The competition is further along on disclosure. OpenAI’s GPT-Live-1 reports 0.798 s turn-taking latency on Full-Duplex-Bench v1 and is available through the API and ChatGPT. NVIDIA’s NemotronLabs VoiceChat 11B is an open-weights model (OpenMDW 1.1) with 448 ms latency and 82.5% tool selection on FDB v3. It runs on A100 through B200 GPUs and is on Hugging Face. Block’s Dessa trains on human speech only, with no text, and runs on a high-end MacBook Pro. Those benchmark versions differ, so the numbers aren’t directly comparable. Still, they show what a serious launch looks like.
Tofu-1 offers little of that. Parameter count, latency, tool-use scores and pricing are all undisclosed. The weights are closed. The only figure is self-reported: over 92% of listeners could not spot it as AI in screenings. Bland says it trained on millions of hours of real conversation, which may be its actual edge, since call-center-grade data is hard to get.
Access is limited to a small number of enterprises, with banking, healthcare, insurance and government named as targets. The sign-up form asks for monthly call volume, from under 1,000 to 50,000+. Bland cites January for broader availability.
Does it matter? Potentially. Regulated industries want voice agents that handle interruptions and noise without sounding robotic, and cascades struggle there. If Tofu-1 holds up, it competes directly with OpenAI’s realtime stack for enterprise phone work. But until Bland publishes latency, tool-use results and pricing, treat it as a credible direction, not a proven product. Developers should wait for numbers before rebuilding anything.



