Wink Pings

Bland Unveils Tofu-1: A Voice Model That Can Interrupt Mid-Call, No Public Performance Report Released

Bland has released Tofu-1, a full-duplex speech-to-speech model targeted at noisy, high-stakes telephone call scenarios. The product is currently in the enterprise research preview phase, with no public disclosures of its benchmarks, model scale, or pricing.

Bland has officially launched Tofu-1. As reported by MarkTechPost author Michal Sutter, this is a full-duplex speech-to-speech model built for telephone scenarios, with core optimizations for noisy environments and high-stakes calls. It is now in the enterprise research preview stage, with no public benchmarks, disclosed model size, or announced pricing available.

The term "full-duplex" carries a very specific meaning in the voice interaction field: the model can listen and speak at the same time, handling interruptions, mid-speech interjections, and overlapping dialogue. Traditional phone bots work much like walkie-talkies, waiting for users to finish speaking before giving a response. In real-world calls however, users often make corrections mid-sentence, background chatter can break in, and phone signals may fluctuate unexpectedly. Telephone calls are also high-cost scenarios: billing inquiries, appointment scheduling, after-sales support, debt collection, any misheard line may directly generate a formal service ticket. Tofu-1 is precisely positioned to solve all these common pain points.

The core issue is that there is still a gap between its claimed positioning and verifiable real-world performance, as key supporting data is not released. Without public benchmarks, developers cannot compare its latency, word error rate, interruption recovery performance, and noise robustness. Without disclosed model scale, it is impossible to estimate inference costs. Without clear pricing, enterprises cannot carry out procurement budget calculation. The enterprise research preview is a common product launch path: developers partner with a small group of clients to run real user traffic first, before deciding whether to disclose full technical details. For Bland, this method effectively controls operational risks; for the public, there is nothing to reference other than the official product statement for now.

The most striking information in the press release is actually all the missing data points. Tofu-1 is worth attention, because full-duplex telephone speech is a technically challenging yet extremely valuable direction. But it is still far from being fully evaluable. We will need to wait for the release of its benchmarks, model scale, pricing and real call samples, before we can judge whether it has properly mastered the handling of mid-speech interruption, noise and low latency.

Original link: MarkTechPost

发布时间: 2026-11-15 18:06