Bland introduces Tofu-1, a speech-to-speech model for hard phone calls

Bland introduced Tofu-1, its first speech-to-speech model. It's full duplex and built for high-stakes phone calls in bad conditions, like moving cars, hospital rooms, weak signal and several people talking at once.

Tofu-1 processes audio directly instead of transcribing it first. Bland says it tells speakers apart and works out who a phrase is meant for, separates voices from noise such as a TV, a crying child, traffic, an intercom or keypad tones, and hears hesitation, frustration and urgency. Callers can interrupt or correct it mid-sentence.

It keeps talking while it works on a task, such as an identity check, a claim status lookup or a payment, and doesn't say the task is done until it is. It can also generate a whole call scene, with a second speaker, room sound and line noise. Bland says the audio in its launch video was fully generated by Tofu-1 and isn't a real call.

You can't sign up and use it yet. Tofu-1 is a research preview that a limited number of enterprises in banking, healthcare, insurance and government can request. There's no self-serve endpoint, and Bland hasn't published benchmarks. Bland describes Tofu-1 as the first of a family of speech-to-speech models.