Skip to main content
v2026.11,986 entries · CC-BY 4.0
CASRAIAsk CASRAI · AI assistant for researchAsk a research question. Answers come with their sources.Funder rules, compliance, methods and publishing. Answers link the sources they rest on.

Editorial · CASRAI · AI and ML research outputs

Bland AI Unveils Tofu-1 Full-Duplex Speech-to-Speech Model in Research Preview

Bland AI unveiled Tofu-1, a full-duplex speech-to-speech model designed to listen, reason, and speak simultaneously across noisy telephony channels.

Published 16 Nov 2026· 7 minute read

Ask CASRAI · free to try

Ask about this story

2 free questions a day, no account.

An AI assistant for research: funder rules, compliance, methods and publishing. It links the sources behind its answers and labels answers that come from a web search.

Answers draw on CASRAI's guides and dictionary plus the federal and funder documents we index: Federal Register, Grants.gov, Regulations.gov and UKRI.

Works on this site and inside Claude, Cursor and the AI tools you already use.

Everything CASRAI publishes — this page, the dictionary, the guides and the news — stays open to read, with no account and no card.

CASRAI is the reference for research administration — bookmark it for the next question.

On November 15, 2026, San Francisco voice artificial intelligence company Bland AI unveiled Tofu-1 in an enterprise research preview, introducing a native full-duplex speech-to-speech model engineered to listen, reason, and speak simultaneously across degraded real-world audio channels without relying on intermediary text transcripts. Designed specifically for chaotic telephony conditions such as emergency dispatches, moving vehicles, and noisy hospital facilities, the system marks a major operational departure from traditional multi-stage voice cascades for researchers and enterprise practitioners.

What Happened: The Transition to Direct Speech-to-Speech Architecture

Conversational voice agents have historically operated via a sequential cascade architecture. When an individual speaks into a telephone line, an automatic speech recognition (ASR) engine converts acoustic waves into text, a large language model (LLM) generates a textual answer, and a text-to-speech (TTS) synthesizer renders that output back into sound. While modular, this three-step pipeline introduces structural latency—frequently exceeding 1,500 milliseconds—and discards paralinguistic nuance, including emotional tone, pacing, breathing, hesitation, and acoustic spatialization.

Bland AI announced that Tofu-1 eliminates this intermediate text abstraction entirely. According to technical documentation published on the Bland AI official research platform, Tofu-1 functions as an end-to-end neural model that processes raw audio directly into audio representations. By maintaining continuous auditory modeling, the system retains conversational markers that are routinely flattened by text tokenization, allowing the model to adapt dynamically to the human caller’s conversational cadence.

Full-Duplex Interaction and Concurrent Reasoning While Speaking

A primary bottleneck in automated telephony is conversational turn-taking. Conventional voice bots function in half-duplex mode: they require the caller to fall completely silent before beginning inference, and any background sound or mid-sentence correction can cause awkward stuttering, abrupt cutoffs, or conversational failure. Tofu-1 implements full-duplex interaction, maintaining active listening channels while synthesizing its own voice.

Crucially, the system introduces a distinction between conversational backchannels and substantive interruptions:

  • Backchannel discrimination: When a caller utters brief affirmative acknowledgments such as “right” or “mm-hmm” while the model is delivering information, Tofu-1 recognizes that conversational floor remains with the agent and continues speaking without pause.
  • Mid-sentence corrective steering: If a caller interrupts to correct critical information—such as amending an address or account number—Tofu-1 catches the correction mid-syllable, halts its planned output, and pivots its reasoning to incorporate the amended input.
  • Thinking while talking: During enterprise workflows like checking claims status, looking up clinical records, or processing identity verification, Tofu-1 maintains conversational engagement rather than imposing dead air, updating its internal operations if the user amends details before the transaction finalizes.

In product demonstrations highlighted by MarkTechPost, the company showed the model navigating multi-step database queries while speaking with a customer, refusing to confirm completed transactions until backend API confirmations were finalized.

Acoustic Noise Separation and the TV Test

Telephony audio is notoriously degraded, compressed over low-bandwidth codecs at sample rates of 8 kHz to 16 kHz. Furthermore, calls frequently originate from hostile acoustic spaces filled with crying infants, highway traffic, blaring radios, and hospital overhead pages. While traditional noise-cancellation filters attempt to strip non-speech frequencies, they fail when background sound consists of intelligible human speech from an unaddressed source.

To address this, Bland AI formulated what it terms the “TV Test.” In a typical failure scenario for cascade bots, an in-room television broadcast utters the word “Yes” or “Call now” while an agent asks a confirmation question. Cascade systems transcribe the word “Yes” and erroneously execute the transaction. Tofu-1 uses spatial and prosodic tracking to separate the primary caller from ambient audio streams. The model attributes speech to its actual source and verifies whether an utterance was directed at the agent before updating conversational state.

Additionally, Bland claims Tofu-1 can synthesize complete environmental audio scenes in a single inference pass, generating multi-speaker interactions, room acoustics, and line characteristics simultaneously. As reported by Glitchwire, Bland reported that in blind screening evaluations, more than 92 percent of human evaluators could not distinguish Tofu-1’s synthetic telephone audio from genuine acoustic recordings.

Architectural Comparison: Closed Enterprise Models vs. Open-Weight Systems

The unveiling of Tofu-1 highlights an escalating architectural divergence between proprietary managed voice platforms and open-weight academic research frameworks. On November 11, 2026, the academic research consortium OpenMOSS released its own full-duplex system, documented in CASRAI’s analysis of the MOSS-Live interaction system. While both paradigms pursue concurrent listening and speaking, their operational design and accessibility differ sharply.

Dimension Bland AI Tofu-1 MOSS-Live (OpenMOSS) Conventional Cascade Voice Bots
Architecture Native end-to-end speech-to-speech Decoupled streaming multimodal foundation model Sequential pipeline (ASR to LLM to TTS)
Weights Access Proprietary closed-weights API Open-source weights and codebase Varies by underlying third-party models
Target Infrastructure Enterprise telephony networks (PSTN/VoIP) Local GPU clusters and web streaming interfaces Cloud API orchestrators and webhooks
Audio Domain Focus 8 kHz telephony, speaker separation, TV Test Omnidirectional conversation, video, multimodal Clean, near-field microphone speech
Task Concurrency Integrated “thinking while talking” state loop Decoupled background reasoning and tool workers Linear, turn-gated synchronous execution

Bland AI CEO Isaiah Granet has stated that building custom speech models from the ground up is necessary because off-the-shelf foundation models cannot reliably handle the messy edge cases of live phone calls. Supported by a 50 million dollar Series C financing round in June 2026 led by Dell Technologies Capital—bringing aggregate funding above 100 million dollars—Bland currently processes over 3.5 million calls weekly across healthcare, insurance, and financial institutions.

Implications for University Researchers, Research Offices, and Clinical Governance

The emergence of end-to-end speech-to-speech agents carrying near-human acoustic naturalness presents immediate governance, regulatory, and methodological challenges for academic institutions and clinical research facilities.

First, research ethics boards and Institutional Review Boards (IRBs) face pressing consent challenges. When universities or affiliated teaching hospitals deploy automated voice systems for patient intake, clinical trial participant recruitment, or post-operative follow-up calls, subjects may struggle to discern whether they are conversing with a human practitioner or an artificial agent. Institutional policies must require explicit, automated disclosure prior to interaction, adhering to statutory standards outlined in CASRAI’s framework for telehealth informed consent. The legal and ethical doctrine of informed consent mandates that research subjects understand the autonomous nature of data collection, potential transmission paths, and storage parameters.

Second, the rise of closed-weights voice models creates severe reproducibility deficits in speech science. In accordance with established norms of research integrity, scientific claims regarding model robustness, bias, and acoustic fairness require verifiable evaluation methodologies. Because Bland AI has not released model weights, training corpora splits, or parameter counts, university laboratories cannot independently audit Tofu-1’s error rates across accented speakers, pediatric speech, or cognitive speech impairments. Scientific progress in speech interaction depends on transparent open data practices, including publicly accessible acoustic stress-test benchmarks like those championed in open-weight initiatives.

Third, research security officers and IT administrators must confront authentication vulnerabilities. The demonstrated ability of speech-to-speech models to generate multi-speaker dialogue, realistic background ambiance, and emotional prosody undermines voice-biometric authentication protocols commonly used by university financial offices and research hospital patient portals. Independent tech security analysts have cautioned that acoustic scene generation tools require strict anti-spoofing verification before integration into institutional infrastructure.

Industry Reaction and Critical Verification Gaps

Market analysts and technical observers have reacted with a mix of interest and skepticism. Enterprise technology reviews featured on Substack AI industry analyses emphasized that Tofu-1 represents Bland’s most ambitious model family launch, positioning the startup directly against hyperscaler voice offerings from OpenAI and Google. Furthermore, brief video snippets published via the Bland AI official demonstration channel highlight the model’s seamless handling of background street noise while discussing billing statements.

However, industry analysts have emphasized that verification remains thin. Bland has not published peer-reviewed ablation studies or standardized Word Error Rate (WER) and latency metrics under standardized telecommunication test conditions. Without independent audit data, enterprise claims regarding noise separation and speaker tracking cannot yet be considered scientifically verified capabilities.

What Happens Next: Rollout Schedule and Governance Next Steps

Bland AI has opened request submissions for its enterprise research preview, focusing initially on institutional partners in healthcare, banking, insurance, and municipal administration. The company intends to test operational edge cases in production before determining a general availability release date and commercial pricing structure.

Over the coming months, telecommunications regulators and academic compliance committees are expected to scrutinize full-duplex speech-to-speech systems more aggressively. Research institutions planning to utilize speech-to-speech agents in clinical trials or human-subjects research must update their IRB protocol templates, verify state-level wiretapping and two-party consent statutes, and ensure robust cryptographic auditing of all audio data streams.

Sources

Referenced across the research world

University of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logoUniversity of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logo
  • University of Cambridge logo
  • Columbia University logo
  • Crossref logo
  • University of Edinburgh logo
  • Harvard University logo
  • University of Oxford logo
  • Princeton University logo
  • Stanford School of Medicine logo
  • University College London logo
  • ORCID logo

View CASRAI adoption →