Block Unveils Dessa: Full-Duplex Speech Model Trained on Human Audio, Not Text

Block, the payments and financial services company behind Square and Cash App, unveiled Dessa on October 23 — a speech foundation model that its AI research lab says it trained from scratch on recorded human conversation, with no text data anywhere in the pipeline. There is no transcription step, no pretrained language model underneath, and no conversion of audio to words at any point during training or inference. The model takes audio in and puts audio out, and Block says everything it knows reached it as sound.

That is a straightforward claim with significant technical implications. Nearly every voice AI product deployed at scale today — from OpenAI's real-time voice experience to Google's Gemini Live to enterprise call-center systems — routes audio through a text language model as its reasoning layer. Speech recognition converts what you say into tokens, the language model generates a response in text, and text-to-speech synthesis reads it back. Dessa is Block's argument that this three-stage pipeline is not architecturally required and that its conversational information cost is larger than the field has acknowledged.

Block's project page states the claim directly: "As far as we know, Dessa is the world's first AI you can have a realistic live conversation with that was trained only from human speech." The hedge is deliberate and appropriate. The company is not asserting that no prior research touched this space — it is asserting that no prior deployed, conversational model with realistic speech quality has been trained this way. That distinction matters for what comes next.

Speech AI Without a Text Layer

The defining characteristic of Dessa's architecture is full-duplex operation, and it is worth being specific about what that means. Half-duplex voice systems, which include most voice assistants and LLM-backed products, are built around turn detection. The system listens until it determines you have finished speaking, then stops accepting audio, generates a response, and plays it back. You talk; then it talks; then you talk again. The entire pipeline runs sequentially.

Full-duplex means both sides process simultaneously. A full-duplex model continues listening while it is speaking, the way two people on a phone call can overlap each other, interrupt, laugh at the same moment, or trail off mid-thought without the other party freezing until silence is detected. Block says Dessa operates this way: it listens and talks at the same time.

This matters for two reasons. The first is naturalness. Turn detection introduces latency and behavioral artifacts — the pause before a system responds, the way it sometimes cuts you off when you hesitate, the absence of overlapping speech that makes every voice assistant interaction feel like a radio call with a protocol. Full-duplex eliminates those artifacts by design.

The second reason is information. Text strips out a large fraction of what human speech carries. Intonation, rhythm, hesitation, vocal fry, laughter, the slight thinning of a voice under stress — none of this survives transcription to words. A model that never converts speech to text can, in principle, learn what those signals mean from millions of examples of them appearing in natural conversation. Block says Dessa hears and speaks emotion, sarcasm, laughter, a frustrated sigh, and a voice cracking with sadness. The model has no handwritten rules for any of these behaviors; it learned them from the training data.

Two specific conversational behaviors that Block highlights are interruption and backchanneling. Interruption at the right moment — knowing when the other person will welcome being cut off versus when cutting in will feel rude — is a skill that emerges from conversational exposure. Backchanneling is the collection of short acknowledgments ("yeah," "got it," "mm-hm," "ooh") that signal active listening without taking the conversational floor. In natural speech, these appear constantly and are timed with precision. Text pipelines strip them out entirely; they never appear in a transcript because they are not information-carrying in the textual sense. A model trained on raw audio encounters them thousands of times and learns when they are appropriate. Block says Dessa learned both skills from the data rather than from rule-based injection.

Block's demos include clips in which Dessa laughs mid-sentence at something it finds funny and speaks louder when asked to. The company states that nobody steered the model during those recordings. Treating those clips as vendor-selected footage is reasonable; they are not independent evaluation. What they demonstrate is that the target behavior is at least achievable in controlled conditions.

Training on a Single GB200 Rack, Running on a Laptop

Two infrastructure details accompanying the Dessa announcement are striking enough to warrant scrutiny. According to multiple secondary reports — including AI Chat Daily and PANews, which attributed the detail to Block's project page — training Dessa required a single NVIDIA GB200 NVL72 system. Dessa runs on a MacBook Pro with an M4 Max chip at faster-than-real-time inference speed.

The GB200 NVL72 is a rack-scale system that pairs 72 Blackwell B200 GPUs with 36 Grace Arm CPUs in a single NVLink domain. It is not a data center cluster. For context, training runs for large language models in 2024 and 2025 typically consumed thousands of GPUs over weeks or months. A frontier-tier capability trained on one rack would be an unusual efficiency result, though it is not impossible — speech models trained on audio tokens operate in a different regime than text LLMs optimizing over a hundred-trillion-token corpus.

Block does not offer a parameter count for Dessa, which makes it impossible to assess whether the GB200 NVL72 training claim is unsurprising or remarkable. What can be said is that the claim, if it holds, points to a model whose training footprint is accessible to organizations that cannot sustain multi-thousand-GPU clusters. That changes the economics of speech AI research.

The inference story is similarly significant. Apple's M4 Max chip offers up to 128 gigabytes of unified memory and approximately 410 gigabytes per second of memory bandwidth. Running a conversational model at faster-than-real-time on that hardware means latency is not the limiting factor for local deployment. Block's project lead, Ragavan Thurairatnam, has indicated in a post that Dessa can also run on consumer GPUs, though Block has not published hardware requirements for this configuration and that claim has not been independently confirmed.

Block says the model learned language from hundreds of times fewer words than large language models. That claim contains no denominator. A large language model's training corpus might span one to fifteen trillion tokens, depending on the organization. "Hundreds of times fewer" could mean fifty billion or five billion — those are different engineering situations. Without the denominator, the comparative efficiency claim is impossible to evaluate. Block may be making a precise point about conversational data density versus text corpus volume; the arithmetic cannot be checked without the numbers.

How Dessa Differs from Moshi and Every Other Voice AI in Production

The most direct technical comparison for Dessa is Kyutai Moshi, the French AI lab's full-duplex spoken dialogue model unveiled in July 2024. Moshi is important context because it demonstrated full-duplex conversational AI before Dessa and made its weights publicly available under a permissive license. Understanding where Dessa claims to differ from Moshi explains what Block is actually arguing.

Moshi was built on a pretrained text language model called Helium. Kyutai initialized Moshi from Helium's weights, then adapted the model to handle audio by introducing Mimi, an audio codec that converts raw speech into discrete tokens the language model can process. The result is full-duplex — Moshi listens and speaks simultaneously, handles backchanneling, and can interrupt — but its underlying language knowledge came from text pretraining.

Dessa's claimed distinction is that text appears nowhere. The model was not initialized from a text LLM. Its training corpus contained no transcripts. Block says every fact Dessa knows reached it as sound. This is the specific architectural departure from Moshi, and from every other voice AI in production.

OpenAI's real-time voice experience and Google's Gemini Live are text-mediated by design. The reasoning layer is a text language model; speech recognition and speech synthesis are wrappers. This architecture benefits from the entire corpus of text pretraining — factual knowledge, reasoning chains, world history, scientific literature — but processes speech as a derived signal. OpenAI's voice mode is half-duplex, though more responsive than traditional systems; Gemini Live has similar structural properties.

NVIDIA's VoiceChat-11B, mentioned in coverage of the voice AI space, requires an 80-gigabyte GPU for inference — constraining it to professional or enterprise hardware. Sesame has demonstrated conversational voice AI with high emotional naturalness, but its architecture details remain undisclosed. Neither matches Dessa's claimed zero-text training approach.

The competitive shift Dessa represents is not primarily about latency or naturalness, though both matter. It is about dependency. Products built on a text LLM as the reasoning layer inherit that model's pricing, rate limits, update schedule, and data policies. A speech-native model that runs on a laptop would let a developer ship a voice feature with no per-minute cloud bill and no exposure to a third party's model changes. Block has not said whether or when Dessa will be accessible to outside developers; that remains a possibility, not a roadmap item.

There is also a quality-of-signal argument worth stating precisely. When a voice AI routes audio through ASR and TTS, it introduces two additional error surfaces. ASR systems misrecognize words, especially proper nouns, technical terms, and accented speech. TTS systems impose a synthetic quality floor on output that varies by prosody but is rarely indistinguishable from human speech under extended listening. A full-duplex audio-native model removes both error surfaces entirely — what it produces is never passed through a text bottleneck where information is discarded and then partially reconstructed. Whether that architectural advantage materializes in practice depends on whether the model's audio generation quality is itself sufficient, which is what independent evaluation will need to determine.

Read more: SpaceXAI's Grok Voice 3.0 tops the speech-to-speech AI index

The Research Question Block Is Asking

Block frames Dessa's release as a research contribution rather than a product launch, and the framing is intellectually honest. The company is opening a genuinely contested empirical question: how much factual world knowledge can a model acquire from audio alone, and how does that compare to what text-pretrained models learn from written corpora?

This question has a history. Meta AI Research published foundational work on Generative Spoken Language Modeling — GSLM in 2021, demonstrating that models trained on discrete speech units from raw audio without text supervision could learn linguistic structure. The 2021 results showed that word-level patterns, syntactic dependencies, and some semantic relationships could emerge from audio training, but factual recall and open-domain question answering were substantially weaker than text-supervised equivalents. The gap made sense architecturally: written text concentrates information differently than spoken conversation, which is full of reformulations, hedges, and social discourse that carry less explicit factual content per word. Whether the gap has narrowed with better audio tokenization, larger audio corpora, and improved architectures over the five years since that paper is exactly what Block is now positioned to measure.

Block asks three research questions on its project page: how efficiently do speech-trained models learn language and facts compared to text models; how efficiently do they learn new languages; and what kinds of in-context learning they can do from audio context. These are not rhetorical questions. They are the empirical program that justifies a model with no published benchmarks — a model whose capabilities are, by Block's own framing, an open question.

The early-access cohort of AI safety researchers is, in part, a mechanism for generating that evidence. Block is inviting external scrutiny before the company has characterized Dessa's own performance. Whether this reflects genuine epistemic humility, strategic caution, or both, it is an unusual rollout pattern that distinguishes Dessa from typical model announcements built on company-selected benchmark tables.

What Block Has Not Yet Published

Dessa arrives with a meaningful list of undisclosed specifications. There is no published parameter count. No dataset size. No latency figure for end-to-end audio response. No formal benchmark results on any standard evaluation — no word error rate, no automatic speech recognition comparison, no conversational coherence metric, no factual accuracy evaluation against an established dataset. AI Chat Daily confirmed in its coverage that none of these figures have been released.

The verification status of the announcement's key claims is therefore mixed. That Dessa is full-duplex and audio-in/audio-out is stated by Block and consistent across all secondary coverage — this is the architectural framing Block is choosing to make, and there is no reason to doubt it in the abstract. That the model was trained with no text and no pretrained models is a strong claim that is structurally unverifiable from outside the company. The training process is not auditable by independent parties, and the architecture of textless speech training is not uniquely identifiable in a deployed model's outputs.

The training hardware claim — a single GB200 NVL72 — appears in secondary reports attributed to Block's project page. No independent verification of training costs or compute is available. The inference claim — faster-than-real-time on MacBook Pro M4 Max — is the most testable of the hardware claims, and will presumably be verifiable when researchers in the early-access cohort run the model.

Block's characterization of "hundreds of times fewer words than large language models" — a phrase drawn from Block's project page — is a comparison without a denominator. The efficiency framing may be accurate, but the absence of a specific number prevents readers from assessing whether the claim is a meaningful distinction or marketing language.

The one independent evaluative voice available as of October 24 is The Verge's Jay Peters, who described early conversations with Dessa as "a little awkward and uncanny" and compared the model to "an awkward new friend." That characterization is relayed secondhand via Digg's coverage, not from Peters's own article. It is the only non-Block characterization of the model's behavior and should carry appropriate weight — less than an independent benchmark, more than nothing.

Safety Researchers First, General Access Later

Block's decision to open Dessa exclusively to a small group of AI safety researchers before broader release is a direct response to the model's capabilities. The company's own project page states that it wants researchers to look for risks its own team has not anticipated. AI Chat Daily reports that applicants must provide a name and an institutional email from a university, company, or laboratory to qualify. There is no public waitlist with a realistic near-term timeline; the current stage is research preview.

The safety rationale is grounded in a specific capability concern. A speech model that hears and responds to emotional register — that can laugh at the right moment, soften its tone when someone sounds distressed, or modulate volume in response to context — has different manipulation potential than a text chatbot. Text-based AI manipulation is increasingly documented; the analogous capability at conversational audio quality, running in real time, raises the ceiling on what social engineering through AI voice can accomplish.

The existing threat landscape is relevant. Automated voice fraud — calls impersonating family members, bank representatives, or government agencies — is an established criminal category. What changes with full-duplex, emotionally fluent, locally-runnable speech AI is responsiveness and authenticity in ways current recorded or low-latency synthetic audio does not achieve. Block is not the first to build in this direction, but it appears to be the first to build without a text backbone and to frame the safety stakes in those specific terms.

Researchers who gain access will likely probe whether Dessa can be steered into deceptive intimacy, how it handles conversations involving distress or crisis, what conversational information it retains between sessions, and whether its audio generation can be extracted and repurposed for impersonation. The enrollment requirement — institutional email from a university, company, or lab — creates a light accountability layer, though it will not constrain anyone with institutional credentials who wants to probe the model's failure modes. Block has not published a safety evaluation or red-teaming report alongside the announcement, which means the safety-first framing is currently an intention rather than a documented result. That documentation should be the most pressing deliverable from the early-access phase before any broader rollout.

If the Claims Hold, a Dependency Disappears

The technical question underneath Dessa's announcement is whether a speech model can serve as a complete conversational AI without borrowing its knowledge from text. Block's answer is structured as a research program rather than a product claim — the early-access cohort and the explicit framing of factual grounding as an open question signal that Block does not yet know the full answer either.

If independent researchers confirm that Dessa converses fluently and grounds factual claims accurately from pure-audio training alone, the implications reach beyond voice AI. It would mean the text-LLM stack, which currently underpins nearly every deployed conversational AI system, is not a necessary component of capable speech AI — only a convenient one. The entire design pattern of bolting speech layers onto language models would face an empirically supported alternative.

If the model's factual accuracy proves weak outside curated demo conditions, the announcement remains scientifically interesting — a data point in the longer research program that Meta FAIR began in 2021 — but the product implications narrow. What researchers need to evaluate first is the benchmark gap between pure-audio training and text-initialized models on factual retrieval, not just conversational naturalness. Block has opened the question. The early-access cohort is positioned to answer it.