THE ESSENTIALS IN 30 SECONDS

  • Tofu-1 is a full-duplex speech-to-speech model: it can listen and speak at the same time and keep reasoning while it talks.
  • Bland says it separates the caller's voice from sounds such as televisions, traffic, crying children and hospital intercoms.
  • Access is a research preview for a limited number of enterprises; Bland has published no benchmarks, latency figures or pricing.

San Francisco voice-AI company Bland has unveiled Tofu-1, a speech-to-speech model designed for phone calls placed from cars, hospital wards and busy homes. Glitchwire dates the announcement to 13 November. Bland presents Tofu-1 as the first model in a planned family of speech-to-speech systems and is aiming it at banking, healthcare, insurance and government teams that handle sensitive calls.

What the model is meant to do

Tofu-1 is full duplex, meaning it can hear and speak at the same moment rather than waiting for strict turns. According to Bland's product page, callers can cut in or correct it mid-sentence, and the model is supposed to tell a simple acknowledgment such as "mm-hmm" apart from a correction that should make it stop. It is also designed to keep the conversation going while it works through tasks like identity checks or payment confirmation, and Bland says it will not report a task as finished before it actually is.

The main selling point is noise handling. Bland says Tofu-1 isolates the caller's voice from sounds such as a television, a crying child, road noise, a hospital intercom or keypad tones, and works out which speaker a phrase is addressed to. The company describes a "TV Test," in which the model should follow the caller even when a television in the room is louder. If the caller cannot be understood, the page says the model should ask them to repeat or move somewhere quieter instead of guessing. Tofu-1 can also generate full audio scenes, including a second speaker and background sound, in its output.

What has not been shown

Bland's page contains no benchmarks, accuracy rates or latency measurements, and it labels its timing diagram as conceptual rather than measured. Glitchwire reports a company claim that more than 92% of listeners in closed screenings could not distinguish Tofu-1 from a real recording, but the method has not been published and no independent tests are available. Model size, supported languages and pricing are not disclosed.

Access is limited to a research preview open to a small number of enterprises, which apply through a form on Bland's site; no general availability date is given on the page. Glitchwire also points out that convincing synthetic voices and background audio raise questions about deepfakes and caller verification, an issue that is likely to matter in the regulated sectors Bland is targeting.