📅 November 13, 2026 • bland.ai Tofu-1 is a full-duplex speech-to-speech model from Bland, a San Francisco voice-AI company, released on November 13, 2026 and announced by cofounder and CEO Isaiah Granet on X. It listens while it speaks and reasons while it talks, and opens a planned family of Bland speech-to-speech models. Bland lists four capabilities: understands audio, generates whole scenes, listens while it speaks, thinks while it talks. The first version supports English and Spanish over standard phone lines and follows Bland Speech v3, released on 4 August.
Audio understanding. Bland lists three skills: speaker tracking, which tells speakers apart and who a phrase is addressed to, noise separation across a crying child, television, highway traffic, hospital intercom and keypad tones, and paralinguistics such as hesitation, frustration and urgency. Thinking while talking keeps the conversation moving while the model verifies identity, checks a claim status, looks up benefits or confirms a payment, and reports the result only when the work is done. Bland publishes no model size, no latency, no price and no benchmark scores for Tofu-1.
Whole scenes and the TV Test. Because it models the sound of a call, Tofu-1 can produce a second speaker, overlapping voices, the room around them and line noise, and Bland states every sound in its launch video was generated by Tofu-1. A clear voice from a television is not noise but speech from the wrong source, so Tofu-1 attributes it to the television and asks the caller to confirm, and if the caller cannot be understood it says so and asks for a repeat or a quieter place. Callers can interrupt mid-sentence: a short affirmative means continue, a correction means stop.
The one number and five failures. Per Bland, more than 92 percent of listeners in its internal screenings could not tell Tofu-1 output from a real recording, and the methodology has not been published. Five failures disqualify a model: claiming success when the system failed, committing a wrong identifier, email, spelling or number, acting before a required verification, speaking internal reasoning aloud, and inventing an ability it lacks. The preview sign-up asks for monthly call volume, from under 1,000 to more than 50,000, in banking, healthcare, insurance and government.
IMPACT: OpenAI GPT-Live-1 reports 0.798 s turn-taking latency on FDB v1, NemotronLabs VoiceChat 11B is open-weight with 11B parameters and 448 ms on FDB, and Block Dessa takes human speech only, for safety researchers, vendor figures not directly comparable.
Limits: Per Bland, more than 92 percent is a company claim with no published method, the walkthrough is illustrative, access is case-by-case.
IMPACT: With a cascade losing who spoke, the approval model, not only model quality, decides deployment: a misheard call in a bank becomes an unfiled fraud claim, an unfilled prescription or an unapproved benefit.
Sources:
bland.ai/speech-to-speech
x.com/usebland
marktechpost.com/2026/11/15/bland-tofu-1-full-duplex-speech-to-speech-model
glitchwire.com/news/bland-claims-tofu-1-can-hear-a-caller-through-a-crying-baby-and-a-highway
Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky
