Bland opens a Tofu-1 preview for phone calls with overlapping voices

Bland opened requests for a research preview of Tofu-1 on November 13. The speech-to-speech model is designed to listen while it talks, with noisy business calls as its first test. Bland is recruiting a limited number of organizations in banking, health care, insurance and government. Teams granted access can test how the model handles the interruptions and competing voices they encounter in their call workflows.

A television can sound like a yes

In Bland’s illustrative TV Test, a phone agent asks a customer to confirm that she wants to cancel her insurance policy. She is still looking for a number when the television says yes. A transcript that omits the speaker can make that sound like consent. Bland says its intended response is to check with the caller again. The company labels the scenario an illustration and presents no recorded model result for it.

A television voice can be intelligible even when its words have nothing to do with the call. The same distinction applies when a caller turns aside to speak to someone in the room. Bland says Tofu-1 uses the audio to identify who is speaking and whom they are addressing. When a caller is too hard to hear, the response Bland describes is to ask them to repeat themselves.

Acknowledgments and corrections

Full duplex means the model can take in sound while producing its own speech. A caller can acknowledge part of a number the agent is reading out, or interrupt to correct a digit. Those are two kinds of overlap with different consequences. The acknowledgment can leave the readout going. The correction needs to change it.

Bland’s existing phone platform already gives customers control over interruptions. Its July update added settings for how readily an agent yields to a caller and how long it waits before resuming. The documentation describes applying those settings at individual steps or to incoming calls on a number. With Tofu-1, Bland is describing a model designed to distinguish what an interruption means while it is still talking.

Earlier full-duplex work shows why this distinction matters. In its Moshi paper, the Kyutai lab described tracking two separate audio streams, one for the user and one for the system, so its model could account for overlap. The ability to accept both streams says little on its own about how often a system attributes a particular sentence to the right person. Bland’s illustration of overlapping speech specifically says its segment lengths are conceptual.

Looking up one date while hearing another

Bland gives the example of a caller correcting an appointment date during a lookup. The model is supposed to carry the correction into the work and report the outcome after the lookup. Bland’s existing platform already supports calling external services and transferring callers to a human. The new example concerns what happens to work already underway when the caller changes the request.

Bland posted two audio samples on the research page. One lasts 10 seconds and describes a call from a car while a passenger speaks. The 39-second sample depicts an agent continuing to talk while checking information. In its LinkedIn announcement, Bland said it generated every sound in its launch video, including both voices and the background. The preview page gives no measured error rate for distinguishing callers from other speakers during customer calls.

Bland lists separate usage plans for its existing phone service. The Tofu-1 preview page gives no price and calls for organizations to request access. Its form asks for a work email, phone number, monthly call volume and website. Bland says it will email applicants as access becomes available.