Bland Releases Tofu-1: A Full-Duplex Speech Model for Phone Scenarios

Bland releases Tofu-1, a full-duplex speech-to-speech model for noisy, high-stakes phone scenarios, but technical details remain undisclosed.
Bland has released a full-duplex speech-to-speech model called Tofu-1, designed for noisy, high-stakes real-world phone calls. Unlike the traditional three-stage pipeline of speech recognition, text processing, and speech synthesis, Tofu-1 converts directly at the speech level and supports parallel input-output processing, promising to improve inaccurate turn detection, interruptions, and false silence under half-duplex architectures. The model is currently in an enterprise research preview stage, with no official benchmark results, model size, or pricing disclosed. The article analyzes the value of full-duplex for interaction naturalness in scenarios such as phone customer service, appointments, and debt collection, and notes that the information vacuum makes horizontal comparison difficult; before deployment, enterprises should focus on latency and interruption handling, cost and deployment model, and compliance and observability. Full-duplex voice models are becoming a new focus of competition in voice AI.
The Next Stop for Voice AI: From "You Talk, I Talk" to Real Conversation
Voice assistants in phone communication have long faced an awkward problem: the user is only halfway through speaking, and the assistant has already started answering; or the user is clearly pausing to think, but the system misjudges it as the end of a turn. This "half-duplex" interaction model is essentially closer to a walkie-talkie than to a real conversation.
According to MarkTechPost, Bland has released a full-duplex speech-to-speech model called Tofu-1, specifically aimed at noisy, high-stakes real-world phone call scenarios. Unlike the traditional three-stage pipeline of "speech recognition—text processing—speech synthesis," Tofu-1 performs speech-to-speech conversion directly and supports simultaneous bidirectional processing at the architectural level.
The model is currently in an enterprise research preview stage, and the company has not yet published benchmark results, model size, or pricing information.

What Full-Duplex Actually Solves
The Timing Gap of Half-Duplex
In a typical phone bot architecture, the system must first confirm that the other party has finished speaking before it can start recognition, understanding, and generation—a chain that inherently introduces latency. More troublesome is judging "finished speaking" itself: background noise, accents, and mid-sentence pauses can all interfere with turn detection, leading to interruptions or false silence.
A full-duplex model allows audio input and output streams to be processed in parallel, theoretically maintaining a listening state while the user is speaking, thereby getting closer to the rhythm of human conversation where people "listen, think, and respond at the same time." For phone scenarios such as customer service, appointments, and debt collection, this improvement in interaction naturalness directly affects call experience and task completion rates.
Turn detection is the key technical link for understanding this dilemma. In voice interaction, the system needs to determine whether the user has finished speaking based on acoustic features of the audio stream (such as energy drop, silence duration, and fundamental frequency changes). Such methods are often referred to as VAD (voice activity detection) plus endpointing. Traditional approaches usually set a fixed silence threshold—for example, treating a 500-millisecond silence as the end of a turn. The problem is that natural pauses in human speech vary enormously: thinking pauses can exceed one second, while breath pauses in fast conversation may last only two or three hundred milliseconds. Set the threshold too short, and the system tends to interrupt; set it too long, and responses become sluggish. More advanced approaches introduce semantic and syntactic cues, such as judging whether the current sentence is grammatically complete, but the reliability of such methods drops significantly in noisy phone audio. One of the values of full-duplex models is precisely that they turn turn-taking judgment from an external standalone module into implicit modeling that continues inside the model.
Why Emphasize "Noisy" and "High-Stakes"
Bland deliberately highlights the two qualifiers noisy and high-stakes in its description. Phone audio is usually heavily compressed and contains echo and background voices, making it very different from quiet studio speech recognition test sets; high-stakes means the model has extremely low tolerance for incorrect replies—mishearing an appointment time or misreading an amount can cause real losses. This suggests that Tofu-1's design goal is not a general-purpose voice assistant, but a production tool embedded directly into enterprise call workflows.
Behind the Information Vacuum: Business as Usual for a Research Preview
One detail: this release provided almost no verifiable technical details. No parameter count, training corpus size, or latency metrics were disclosed, nor was there any horizontal comparison with similar full-duplex models.
This "release first, disclose later" approach is not uncommon in the voice model field. Enterprise research previews are typically used to collect real-world feedback and lock in early customers while avoiding premature disclosure of data that could invite unfavorable comparisons. For developers following this field, this means that at present it is impossible to judge whether Tofu-1 truly outperforms existing solutions in noisy phone scenarios; they can only wait for a subsequent technical report or third-party testing.
From an industry perspective, full-duplex voice models are becoming a new focus of competition in voice AI, and multiple vendors are exploring speech-native architectures to replace cascaded pipelines. Tofu-1's entry shows that this route is moving from the lab toward enterprise deployment, but the lack of public evaluation also makes it impossible to say for now "who is ahead."
Three Things Enterprises Should Watch Before Deployment
Latency and interruption handling. Full-duplex only means parallel support at the architectural level; the actual experience depends on the model's ability to separate overlapping speech and its response timing. In phone scenarios, users often interject while the system is speaking. Whether the model can handle this overlap smoothly is key to usability.
Cost and deployment model. Pricing has not been announced, but speech-to-speech models usually have high compute requirements. Enterprises need to assess whether to use a cloud API or local deployment, and whether the cost per unit of call duration is controllable.
Compliance and observability. Phone calls involve recording, privacy, and industry regulation. The extent to which the model is auditable and whether it can output call records and decision rationale will directly affect adoption in high-stakes industries such as finance and healthcare.
Conclusion
Tofu-1's release pushes competition in full-duplex speech-to-speech models from conceptual discussion to the enterprise preview stage, and the direction is worth watching. But until the three pieces of the puzzle—benchmark data, model size, and pricing—are filled in, it is more like an admission ticket than a report card. For enterprises currently evaluating phone voice solutions, staying attentive and waiting for reproducible evaluation results is more prudent than rushing to replace existing systems.
Background Supplement
The core difference between speech-to-speech models and traditional cascaded pipelines lies in the information pathway. In a cascaded approach, audio is first transcribed into text, the text is then handed to a language model to generate reply text, and finally TTS synthesizes speech; each conversion loses information such as tone, emotion, and pause rhythm, and transcription errors are amplified step by step along the chain. Speech-to-speech models instead perform understanding and generation directly in the audio representation space, usually using discrete speech units or continuous acoustic encodings as intermediate representations. The benefit is that paralinguistic information is preserved and end-to-end latency is lower; the cost is that training data is harder to obtain, controllability is worse, and intermediate results are harder to inspect directly than text. For phone scenarios, paralinguistic information (such as hesitation or impatience) is often as important as semantics, which is one reason speech-native architectures are favored.
Full-duplex is a term borrowed from communications, originally meaning that both parties can send and receive signals at the same time, as opposed to half-duplex (only one direction can transmit at any given moment). In voice models, full-duplex means the model continuously receives and processes the input audio stream while outputting audio, rather than resuming listening only after output ends. In engineering terms, this requires the model to handle overlapping speech—the so-called "self-speech cancellation" or echo cancellation problem—where the model needs to distinguish which sounds come from the user and which come from the speech it has just generated; otherwise it may fall into self-oscillation or false responses. In addition, parallel processing places higher demands on inference scheduling and memory bandwidth, which is why full-duplex models generally have relatively high computational costs. The real difficulty is not allowing parallelism at the architectural level, but whether the model can make choices consistent with human conversational habits at moments of overlap—for example, stopping speaking in time to let the user interject.
One-line Summary
Bland releases Tofu-1, a full-duplex speech-to-speech model for noisy, high-stakes phone scenarios, but technical details remain undisclosed.
Paragraph Summary
Bland has released a full-duplex speech-to-speech model called Tofu-1, designed for noisy, high-stakes real-world phone calls. Unlike the traditional three-stage pipeline of speech recognition, text processing, and speech synthesis, Tofu-1 converts directly at the speech level and supports parallel input-output processing, promising to improve inaccurate turn detection, interruptions, and false silence under half-duplex architectures. The model is currently in an enterprise research preview stage, with no official benchmark results, model size, or pricing disclosed. The article analyzes the value of full-duplex for interaction naturalness in scenarios such as phone customer service, appointments, and debt collection, and notes that the information vacuum makes horizontal comparison difficult; before deployment, enterprises should focus on latency and interruption handling, cost and deployment model, and compliance and observability. Full-duplex voice models are becoming a new focus of competition in voice AI.
Related articles

I Built an AI Skill That Edits Videos End to End
A YouTuber turned video post-production into an AI skill: automatic denoising, audio optimization, retake removal, transcription, subtitles, plus YouTube titles/descriptions and Instagram captions and covers. This article breaks down the seven-step human-AI division of labor and how to replicate it.

Anthropic's Always-On Agent Design Guide: Breaking Down the Five Core Modules
Anthropic's *Building always-on agents* design guide breaks always-on agents into five modules: event inbox, timers, memory, permissions, and outbox. Here's a module-by-module breakdown of their roles and engineering trade-offs, plus why developer feedback on Reddit points to memory as the weakest link.

The AI Era: A Recursive Starting Point
Starting from a popular Reddit thread title, this article explores the recursive loop paradigm in AI development, analyzing self-reference in data, code, and evaluation, and discussing consensus and divergence across perspectives.