Block unveils Dessa, a full-duplex speech-to-speech model trained only on human conversation
Block, the payments and bitcoin company co-founded by Jack Dorsey, has introduced Dessa, a research preview of a conversational voice model that takes audio in and produces audio out without passing through text at any stage. According to the company's launch page, the model was trained from scratch on recordings of human conversation alone, with no written text and no pretrained language model underneath it, and it operates in full-duplex mode, which means it keeps listening while it is speaking and can decide on its own when to interrupt, when to wait and when to offer short acknowledgements such as "yeah" or "got it".
The announcement was made on October 23, 2026, through a dedicated page on Block's website and a post on X by Ragavan Thurairatnam, who leads the project. In that post he described Dessa as the first AI model a person can have a realistic conversation with that was trained entirely on speech, said it runs on consumer hardware and called the release "a GPT3 moment for speech". Block's own page is more measured on the first claim, stating that Dessa is, as far as the company knows, the first AI able to hold a realistic live conversation after learning only from human speech.
The technical disclosures are short but specific. Block says training ran on a single NVIDIA GB200 NVL72 system, that inference runs faster than real time, and that the model runs locally on a MacBook Pro with an M4 Max chip or better. The company also says Dessa learned language from "hundreds of times fewer words" than large language models need. Access is deliberately narrow: a small group of AI safety researchers is receiving early access, a public waitlist is open, and no weights, API, pricing, parameter count, benchmark results or technical paper have been published.
··········
HOW A SPEECH-ONLY CONVERSATION MODEL DIFFERS FROM TODAY'S VOICE ASSISTANTS.
Most commercial voice assistants are still built as cascades. A speech recognizer converts the user's audio into text, a text language model writes a reply, and a speech synthesizer reads that reply aloud. The approach is modular and easy to improve one component at a time, but every stage boundary discards information. Transcription removes timing, hesitation, laughter and tone of voice, and the synthesizer has to reconstruct prosody from text that never contained it. Turn-taking is usually handled by a separate voice activity detector that waits for silence before handing control to the model, which is why many assistants either interrupt users mid-thought or leave noticeable pauses.
End-to-end speech models try to remove those boundaries by modelling audio directly, typically after converting the waveform into discrete acoustic tokens with a neural codec. Research in this area has followed two broad paths. Textless spoken language modelling, such as Meta's generative spoken dialogue work published in 2022, learned dialogue behaviour directly from two-channel conversational audio without any text supervision, but those systems remained limited in linguistic coherence. More recent full-duplex systems, including Kyutai's Moshi, reached far more fluent conversation by building on a pretrained text language model and predicting text tokens alongside audio as an internal scaffold.
Dessa's distinguishing claim sits between those two lines of work. Block describes a full-duplex conversational model that reaches realistic dialogue without borrowing linguistic knowledge from a text model at all. If that holds up under outside evaluation, it would show that conversational competence, including vocabulary, grammar and pragmatic behaviour such as backchanneling, can be acquired from listening at a scale Block describes as far smaller in word count than text pretraining. The company frames this in developmental terms on its launch page: people learn to talk by listening, and Dessa was trained the same way.
··········
WHAT BLOCK HAS DISCLOSED ABOUT TRAINING AND INFERENCE.
The most concrete figure Block provides is the compute footprint. A GB200 NVL72 is a single rack-scale system that combines 72 NVIDIA Blackwell GPUs and 36 Grace CPUs in one NVLink domain. That is a substantial machine, but it is a single rack rather than the multi-thousand-GPU clusters typically associated with frontier model training, and Block emphasizes the word "one" in its description. The company does not say how long training took, how many hours of audio were used, where the recordings came from or how the conversational data was licensed or consented.
Block also explains why it considers the data-efficiency claim notable. Its page observes that a single spoken word takes thousands of times more bytes than the same word written as text, so a model trained on speech has to process far more raw data per unit of linguistic content than a text model does. Learning language from hundreds of times fewer words, by Block's account, is therefore the counterweight to a much heavier input signal. Neither ratio is quantified beyond those orders of magnitude, and the comparison baseline, meaning which language models and which token counts, is not specified.
On the inference side, the claim that Dessa runs faster than real time on an Apple laptop implies a model small enough to fit in the unified memory of a high-end consumer machine, although the company has not published a parameter count, quantization scheme or latency figure. Faster-than-real-time generation is a prerequisite for full-duplex operation, because the model must continuously process incoming audio while producing its own output stream without falling behind the conversation.
........
Item | Status according to Block |
Input and output | Audio in, audio out |
Conversation mode | Full-duplex, listens while speaking |
Training data | Human conversation audio only |
Text or pretrained models | None used |
Training hardware | One NVIDIA GB200 NVL72 |
Local inference | MacBook Pro M4 Max or better |
Inference speed | Faster than real time |
Parameters and latency | Not disclosed |
Benchmarks or paper | Not published |
Weights and API | Not available |
........
··········
THE DEMONSTRATIONS AND WHAT THEY CAN AND CANNOT SHOW.
Block's page presents several recorded conversations and states that every clip is a live recording of a person talking to Dessa, with nobody steering the model and subtitles added afterwards. Each clip is provided with two synchronized speaker tracks. In one, Dessa laughs in the middle of a sentence and raises its volume when asked to speak up. In another, it completes a speaker's sentence after the person blanks on a movie title. A third shows it gasping and responding with what the page describes as teary gratitude to news of a surprise anniversary. Other clips are presented as examples of wide variation in intonation and speed and of empathetic advice.
The company says Dessa chooses its own voice and personality and decides what to say, when to say it and how. One clip pairs Dessa with GPT-Live-1, a voice model from another developer, and the page notes that both voices are AI-generated and that the other developer is not affiliated with Block and has not reviewed or endorsed Dessa.
Recorded demonstrations are useful evidence that the behaviour exists, but they are selected by the developer and cannot establish how often the model performs this way, how it handles noisy audio, accents or multiple languages, or how factually reliable its spoken answers are. Block's page does not discuss hallucination, factual accuracy, language coverage or failure cases. Those questions matter more for a speech-only model than for a cascaded system, because there is no intermediate transcript that an operator can log, filter or audit before the response is spoken.
··········
WHY BLOCK SAYS IT BUILT DESSA AND HOW IT IS HANDLING ACCESS.
Block lists three motivations on the launch page: curiosity, better understanding between humans and AI, and safety. The safety argument centres on emotional perception. According to the company, current AI is still nearly deaf to the emotion in people's voices, and it gives the example of a voice assistant that carries on normally while the person speaking is in serious pain. A model that learns directly from audio, rather than from transcripts, can in principle pick up on cues such as strain, hesitation or distress that text discards.
The same capabilities raise obvious misuse concerns. A model that sounds convincingly human, reacts emotionally and runs locally on a laptop could be attractive for impersonation and voice-phishing if it were widely distributed. Block's response so far is restricted access. The page says the company is giving a small group of AI safety researchers early access in order to find risks it has not anticipated, and it asks interested researchers to get in touch. It does not describe technical safeguards, watermarking, voice-cloning restrictions or the conditions under which the model might be released more broadly. The page also states that launch links are coming soon, without giving a date.
The team behind Dessa is a small research lab inside Block founded by former members of Dessa, the Toronto deep learning startup that Square, now Block, acquired in 2020. The lab says it has built deep learning models since 2012 and generative speech models for more than eight years, and that it reports directly to Jack Dorsey. The page lists seven researchers, including Thurairatnam as lead, alongside current and former interns.
··········
TECHNICAL ASSESSMENT OF THE DESSA RESEARCH PREVIEW.
What is established is limited to Block's own statements and recordings. The company says it has trained a full-duplex speech-to-speech conversation model from scratch on human conversation audio, without text or pretrained language models, using a single GB200 NVL72 rack, and that the result runs faster than real time on a high-end MacBook Pro. The demonstrations show natural turn-taking, backchanneling, laughter and expressive prosody of a kind that cascaded voice assistants generally struggle to reproduce.
What remains unverified is almost everything needed to place Dessa against other systems. There is no parameter count, no description of the audio tokenizer or architecture, no training-data size or provenance, no latency measurement, no benchmark on spoken question answering or dialogue quality, and no independent evaluation. The claim that Dessa learned language from hundreds of times fewer words than LLMs cannot be checked without knowing the size of its corpus and the comparison point.
The release is best read as a research claim with strong qualitative demonstrations rather than a product. If outside researchers confirm that coherent, full-duplex conversation can be learned from speech alone with single-rack compute, it would challenge the assumption that capable voice models need to be built on top of text language models. Until Block publishes technical details or opens broader access, that conclusion rests on the company's account.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org
