Bland Launches Tofu-1, a Full-Duplex Speech-to-Speech Model for Calls

Why it matters

Bland, a San Francisco voice-AI company, has released Tofu-1, a full-duplex speech-to-speech model, in a limited research preview for enterprises. Unlike cascade voice agents, which convert speech to text before processing, Tofu-1 works with audio directly, so tone, hesitation and background sound remain available to the model. Bland says it can track speakers, separate voices from noise such as traffic, hospital intercoms or keypad tones, and detect caller frustration or urgency. Callers can interrupt it mid-sentence, and it reports results only when a task is genuinely complete. The model is aimed at difficult call environments, including moving cars, hospital rooms and weak-signal lines. Bland names banking, healthcare, insurance and government as target sectors. The company has not published the model's architecture, size, latency, pricing or independent benchmarks, and it acknowledges that its own demonstration materials are illustrative rather than recorded model output. Competitor comparisons are likewise vendor-reported and not directly comparable. The release signals a broader industry shift toward speech-native voice agents, though claims remain largely unverified.