Bland presents Tofu‑1 as a full-duplex speech-to-speech model: a system that processes audio and can listen while speaking. Its announcement dates to November 13, 2026; the project also appears in the verified November 16 GenAI Works newsletter. Access is a limited enterprise research preview, with applications and subsequent admission. This is not a verified generally available API or downloadable model.
Who is speaking matters as much as the words
Bland’s page emphasizes distinguishing callers from background voices and handling corrections during a conversation. The company provides audio samples. It also labels the television scenario and timing graphics as illustrative rather than measured outcomes. Those qualifications limit what can be inferred from the presentation.
Our editorial reading is that recognizing words and having authority to act on them are separate problems for a telephone agent. Even a perfect transcript would not establish that a particular “yes” was addressed to the system. When another person, a television or an interruption enters the scene, the application needs to preserve the boundary between audible speech and confirmed intent.
Consider changing an appointment. The relevant test is not simply whether a new date appears in the transcript. A reviewer should check that the old selection was superseded, the current confirmation was used and the final record reflects it correctly. This is a proposed evaluation scenario, not a Tofu‑1 result we observed.
Moving from conversation to action
An agent can sound fluent while waiting for an external service. Evaluation should therefore distinguish a statement that work is in progress from a statement that an operation has completed. A convincing spoken “done” is not evidence that a change was actually recorded. Confirmation needs to come from the system responsible for the operation.
For a potential pilot, SIExpo would look for calls with a clearly defined correct outcome: a corrected number, a declined action, conflicting voices and a dropped connection. It would be useful to count safe stops when intent remains unclear, alongside successful completions. This is not a proposal to use real sensitive conversations without an appropriate research and operational setup.
A further useful distinction is between a model’s conversational behavior and the surrounding application’s permissions. A voice component might correctly hear a request while the application still lacks authorization to perform it. Testing the combined workflow is therefore more informative than judging the pleasantness of a sample recording.
The checked announcement does not justify promises about a particular latency, Bulgarian-language support or reliability across every kind of telephone connection. SIExpo has not conducted independent audio testing. The development concerns an architectural direction and applications for a restricted preview; practical value still needs to be established through actual tasks and verifiable final actions.
Editorial verification: 2026-11-16T17:16:38.474339+00:00.



