Bland has released Tofu-1, a full-duplex speech-to-speech model in research preview designed to handle phone calls from noisy environments such as moving vehicles, hospital rooms, and weak-signal lines. Unlike cascade voice agents that convert speech to text, process it with a language model, and synthesize speech back, Tofu-1 processes audio directly, preserving tone, hesitation, background sounds, and paralinguistic cues like frustration or urgency. The model features three core audio capabilities: speaker tracking to distinguish and attribute speakers, noise separation to isolate voices from sources like highway traffic, TV audio, crying children, hospital intercoms, and keypad tones, and paralinguistic understanding. It supports full-duplex operation, allowing callers to interrupt mid-sentence while the model continues reasoning and keeps the conversation flowing, reporting results only upon task completion. Tofu-1 can also generate realistic call scenes, including secondary speakers, room ambience, and line noise. Bland evaluates robustness via a "TV Test," assessing whether the model follows the caller while ignoring louder TV audio, though the walkthrough is illustrative. Compared to competitors—OpenAI GPT-Live-1, NVIDIA NemotronLabs VoiceChat 11B, and Block Dessa—Tofu-1 does not disclose parameter count, weights (closed), training data specifics beyond "millions of hours of real conversation," hardware, turn-taking latency, or tool-use benchmarks; NVIDIA reports 448 ms latency (FDB) and 82.5% tool selection accuracy (FDB v3). Access is limited to enterprise research preview targeting banking, healthcare, insurance, and government sectors, with a sign-up form tiered by monthly call volume. Bland cites over 92% of screeners unable to distinguish the model, with general availability expected in January.

Read original