AboutAI ToolsCursorClaude CodeLovableBolt.newAI tool reviewsTutorialsVibe codingAI agentsMobile apps with AIAll tutorialsBlogSponsorWork With MeJoin 285K+ Subscribers

15 November 2026 · 4 min read

Bland Releases Tofu-1, a Full-Duplex Speech-to-Speech Model Built for Noisy Calls

San Francisco voice-AI company Bland has launched Tofu-1 in research preview, a full-duplex model that listens and speaks simultaneously on phone calls, with direct audio understanding and scene generation.

On 13 November 2026, San Francisco-based voice-AI company Bland released Tofu-1, its first full-duplex speech-to-speech model, in a closed enterprise research preview [1][3]. The model is designed for phone calls in challenging acoustic conditions: moving cars, hospital rooms, and lines with weak signal. It listens while it speaks and reasons while it talks, collapsing the usual cascade of speech-to-text, LLM, and text-to-speech into a single model that handles audio input and output directly [1].

What sets Tofu-1 apart

Tofu-1 moves away from the cascade architecture common in voice agents. Instead of converting speech to text, feeding it to a language model, and synthesising replies from text, the model processes audio natively. That means tone, hesitation, and background sound remain available to the model. Bland has not disclosed the internal architecture [1].

Audio understanding

The product page lists three audio capabilities [1]:

  • Speaker tracking: distinguishes speakers and works out who each one is addressing.
  • Noise separation: isolates voices from a crying child, a TV, highway traffic, a hospital intercom, or keypad tones.
  • Paralinguistics: picks up hesitation, frustration, and urgency.

Full duplex and tool use

Callers can interrupt or correct Tofu-1 mid-sentence. While it works on a request, the model keeps the conversation moving and reports the result only when the work is finished; it never claims a task is done early [1].

Scene generation

Because Tofu-1 models the sound of a call, it can also produce an entire audio environment in a single inference pass – including a second speaker, room acoustics, and line noise. In Bland’s demo video, every sound was generated by the model [3].

The ‘TV Test’

Bland uses a ‘TV Test’ to check whether the model can follow the caller while ignoring speech from another scene, typically louder TV audio. The company describes its walkthrough as illustrative, not recorded model output [1].

How Tofu-1 compares

The table below summarises the limited information Bland has published alongside competitor figures. Numbers are self-reported and not directly comparable because of different benchmark versions [1].

Model Tofu-1 OpenAI GPT‑Live‑1 NVIDIA NemotronLabs VoiceChat Block Dessa
Type Full-duplex speech-to-speech Full-duplex voice model Full-duplex speech-to-speech Full-duplex speech-to-speech
Parameters Not disclosed Not disclosed 11B Not disclosed
Weights Closed Closed Open (OpenMDW 1.1) Not released
Training data Millions of hours of real conversation Not disclosed ~550k hours audio + text Human speech only, no text
Hardware Bland-hosted OpenAI-hosted A100, H100, H200, B100, B200, RTX-6000 MacBook Pro M4 Max or better
Turn-taking latency Not disclosed 0.798 s (FDB v1) 448 ms (FDB) Not disclosed
Tool use Keeps talking while it works Delegates to backend model Live tool calling, 82.5% tool selection (FDB v3) Not disclosed
Access Enterprise research preview API and ChatGPT Hugging Face checkpoint, NVIDIA container Research preview for safety researchers

Access and availability

Tofu-1 is available only to a limited number of enterprises via a research preview. Bland names banking, healthcare, insurance, and government as target sectors. The sign-up form asks for monthly call volume, with options from under 1,000 to over 50,000. Bland has indicated a January availability target [1].

Pricing, model size, and turn-taking latency have not been disclosed. No independent benchmarks exist. The only performance figure self-reported by Bland is that over 92% of listeners could not distinguish Tofu-1 output from a real recording in closed screenings, above; the methodology has not been published [3].

What it means for developers

For developers building voice agents, Tofu-1 signals a push toward single-model architectures that handle the full audio pipeline. The model's claimed ability to track speakers and separate noise could reduce the need for external preprocessing on messy calls. However, the closed weights and proprietary infrastructure mean no fine-tuning or local deployment. Bland runs its own speech recognition, language models, and text-to-speech, and customers cannot swap in external models [3].

The lack of latency figures, parameter counts, and publicly available evaluations makes it difficult to assess where Tofu-1 sits against open alternatives such as NVIDIA’s NemotronLabs VoiceChat, which offers open weights and published latency scores [1]. The scene-generation capability also raises questions about voice deepfakes and caller verification if it works as described. For now, the demo is a company claim worth watching, not a confirmed capability [3].

Sources

  1. Bland Launches Tofu-1, a Full-Duplex Speech-to-Speech Model for Calls
  2. Bland Launches Tofu-1, a Full-Duplex Speech-to-Speech Model for Calls
  3. Bland Claims Tofu-1 Can Hear a Caller Through a Crying Baby and a Highway — Glitchwire
  • tofu-1
  • speech-to-speech
  • full-duplex
  • bland
  • voice-ai

Researched and drafted with AI from the sources above.

Learn to build with AI

Hands-on tutorials on Cursor, Claude Code, AI agents and more.