Bland Introduces Tofu-1: A Full-Duplex Speech-to-Speech AI Model for Challenging Call Environments
San Francisco voice-AI company Bland released Tofu-1, a full-duplex speech-to-speech model for AI phone calls, in research preview on November 15, 2026. Tofu-1 is engineered to handle calls in challenging environments like moving cars, hospital rooms, and weak-signal lines by processing audio input and output within one model, unlike traditional cascade architectures.
Tofu-1's Core Capabilities and Design
Tofu-1 represents Bland's initial offering in a planned series of speech-to-speech models. Its design focuses on maintaining the integrity of conversational nuances often lost in other systems. Unlike cascade models that typically discard elements like tone, hesitation, and background sounds, Tofu-1 retains these signals throughout the audio path. This approach aims to create a more natural and responsive conversational experience.
The model's architecture supports full-duplex communication, enabling users to interrupt or correct the AI mid-sentence. This capability is crucial for dynamic, real-time interactions where immediate feedback is necessary. Bland states that Tofu-1 was trained on millions of hours of real conversation data to achieve its performance.
Key Audio Skills
Tofu-1 incorporates three primary audio processing skills:
- Speaker Tracking: The model can identify and follow individual speakers within a conversation.
- Noise Separation: It can isolate and distinguish speech from various background noises.
- Paralinguistics: Tofu-1 processes non-verbal vocal cues, such as tone and hesitation, which contribute to the overall meaning and context of a conversation.
Additionally, Tofu-1 can generate realistic call scenes, including the presence of a second speaker, ambient room sounds, and simulated line noise, further enhancing its utility for creating immersive conversational environments.
Availability and Hosting Details
Tofu-1 is currently available as a research preview. Bland has confirmed that the model's weights are closed, meaning they are not publicly accessible, and the hosting of the model is managed exclusively by Bland. Details regarding the model's parameter count and internal architecture have not been disclosed.
Comparison with Other Conversational AI Models
The landscape of conversational AI includes various models with different approaches to deployment and functionality. Tofu-1's closed-weight, Bland-managed hosting model contrasts with some other offerings in the market.
- NVIDIA's NemotronLabs VoiceChat 11B: This model ships with open weights (OpenMDW 1.1) and was trained on approximately 550,000 hours of audio combined with text data. Its open-weight nature allows for greater flexibility and customization by developers.
- OpenAI's GPT-Live-1: Similar to Tofu-1, GPT-Live-1 is a closed-source model hosted by OpenAI. While specific details about its architecture and training are not fully public, it operates within a managed environment.
These comparisons highlight the different strategies companies employ regarding model accessibility and deployment, impacting how developers and enterprises can integrate and utilize these advanced conversational AI technologies.
Implications for AI Phone Agents
The introduction of Tofu-1 could have implications for the development and deployment of AI phone agents. Its ability to handle challenging audio environments and maintain conversational fidelity suggests potential applications in customer service, telemarketing, and other sectors that rely heavily on voice communication. The full-duplex capability, allowing for natural interruptions, addresses a common limitation in AI-driven conversations, potentially leading to more fluid and human-like interactions.
Conclusion
Bland's Tofu-1, released on November 15, 2026, as a research preview, marks an advancement in full-duplex speech-to-speech AI for phone calls. By focusing on challenging audio environments and preserving conversational nuances, Tofu-1 aims to enhance the realism and responsiveness of AI-driven voice interactions. Its closed-weight, Bland-managed hosting model positions it differently from open-weight alternatives, offering a distinct option for developers and enterprises exploring advanced AI news in voice technology.
Sources
- GitHub - api-evangelist/bland-ai: Bland AI — independent third-party profile of a public API surface, by API Evangelist. Bland AI is an enterprise-grade conversational voice AI platform for inbound and outbound phone agents. The Bland REST API covers calls, pathways (conversation graphs), voices and voice cloning, personas, tools, knowledge bases, dynamic data, transfer lists, and · GitHub
- socialsent/socialsent/data/lexicons/kuperman.json at master · williamleif/socialsent · GitHub
- Why hiring the weirdos works | TechCrunch
- bland.ai
About the Author

Albert Schaper is a co-founder of Best-AI.org. He focuses on product strategy, AI adoption, practical tool selection, and educational content that helps users compare AI products with clearer context.
More from AlbertWas this article helpful?
Found outdated info or have suggestions? Send us a note.