Beyond the Transcript: How Block’s Speech-Only Local Model Redraws the Boundary Between Voice and Intelligence

On 23 October 2026, Jack Dorsey’s Block released a research preview of Dessa, a full-duplex speech-to-speech model trained from scratch on nothing but recorded human conversation. The lab claims it never saw text, never inherited weights from a pretrained language model, and still learned when to interrupt, when to backchannel with a short “yeah,” and how to carry emotion, sarcasm, laughter, and the small silences that make speech feel alive. It was trained on a single NVIDIA GB200 NVL72 rack and, according to the team, runs faster than real time on a MacBook Pro with an M4 Max or better.

That combination—speech as the sole training signal, full-duplex behavior, and consumer-hardware inference—is unusual enough to force a reconsideration of assumptions that have guided most voice AI work for years. Most production voice systems still treat speech as a temporary encoding of text: audio is transcribed, a language model reasons over the transcript, and a synthesizer speaks the reply. Information that lives only in the sound itself is discarded at the first step. Dessa keeps the sound. The result is not yet a product anyone can download, but the research preview already shows what changes when the model is allowed to learn language the way humans first do: by listening.

What the Lab Actually Released

Block’s AI research group, founded by former members of the Toronto deep-learning startup also named Dessa (acquired by Square in 2020), describes the model as audio-in, audio-out. It is full-duplex, meaning it continues listening while it is speaking rather than waiting for a clean turn boundary. The team says it learned conversational timing from real dialogue, so it produces the short affirmative noises and the mid-sentence repairs that signal mutual understanding. Because it never converts speech to text, the model retains cues that a transcription pipeline would strip: a rising pitch that marks a question, a laugh that softens a disagreement, a pause that indicates hesitation.

Training compute is modest by frontier-model standards. A single GB200 NVL72 contains 72 Blackwell GPUs and 36 Grace CPUs in one liquid-cooled rack. That is substantial hardware, yet it is a single rack rather than the multi-thousand-GPU clusters used for the largest text models. The lab notes that the finished model runs locally on a high-end consumer laptop. No model size, parameter count, or language coverage has been published. There are no public benchmarks, no weights, and no API. Access is limited to a small group of AI-safety researchers who apply with institutional email addresses; a public waitlist is open, and the company has said launch links will appear later.

Official details live on Block’s page at block.xyz/dessa. Independent reporting, including TechCrunch’s coverage, confirms the core claims around training data, hardware, and the absence of text.

Why Skipping Text Changes the Architecture

The conventional speech stack is a pipeline: automatic speech recognition (ASR) produces a transcript, a large language model consumes that transcript, and text-to-speech (TTS) generates the reply. Each stage is optimized separately. The pipeline is modular and benefits from the enormous text corpora that exist for language models, but it also discards information. Prosody, timing, breath, laughter, and overlapping speech are largely lost once the words are written down. Full-duplex behavior is awkward because the system usually waits for silence before it begins to generate a reply.

Dessa collapses the pipeline into a single mapping from audio to audio. The model must discover linguistic structure, conversational norms, and emotional signaling directly from the waveform. The lab argues that this produces more natural turn-taking and preserves the cues that make a conversation feel human rather than transactional. Whether the resulting knowledge is as broad or as reliable as that of a text-trained model remains an open question; the team has not released measurements that would allow a direct comparison. What is clear is that the design choices remove one major source of latency and one major source of information loss.

Running locally further alters the trade-offs. A model that stays on the device does not send audio to a remote server for every turn. That reduces network latency, removes a point of data exposure, and allows the system to function when connectivity is limited. It also constrains model size and complexity to what fits in the memory and compute of a high-end laptop. The claim that inference is faster than real time suggests the team has prioritized responsiveness over maximum capability, at least for the research preview.

Why It Matters Now

Voice has long been described as the natural interface for computing, yet most deployed systems still feel like thin wrappers around text chat. The gap between that promise and daily experience is partly technical: transcription errors, unnatural pauses, loss of tone, and the inability to handle overlapping speech all degrade the interaction. A model that never leaves the audio domain and that can listen while it speaks attacks those failure modes directly.

Privacy and control become practical rather than rhetorical when the model runs on the user’s machine. Conversations that involve sensitive personal information, workplace discussions, or medical details no longer need to leave the device for the model to respond. That property is especially relevant as agents grow more capable and as institutions begin to restrict what third-party agents may do inside their systems. Recent TechMash coverage has examined how banks are already erecting barriers against autonomous agents and how earlier agent incidents illustrated containment challenges; a local speech model does not eliminate those issues, but it changes the surface area of the problem.

The training approach also sits in a broader conversation about data efficiency and modality. Large language models require enormous text corpora. Dessa is described as having learned from far fewer words because it worked directly with conversational audio. If that efficiency claim holds under independent scrutiny, it suggests that high-quality, domain-specific audio data may be more valuable than sheer volume for certain capabilities. At the same time, the absence of published evaluations means the claim cannot yet be weighed against the performance of models that do use text.

Open-weight and local-model trends provide additional context. Meta’s Muse Spark Open was engineered to run on Apple silicon desktops; Moonshot’s Kimi K3.1 Pro narrowed the open-to-closed performance gap on several indexes. Dessa is narrower in scope—it is a speech conversation model rather than a general reasoning system—but it extends the same idea: capable models need not live exclusively in the cloud. For developers building assistants that must feel conversational rather than merely accurate, the local full-duplex approach offers a different set of constraints and opportunities.

What to Watch

Several practical questions will determine whether Dessa remains a research demonstration or becomes a foundation for products. First is independent evaluation. Recorded demos can be selected; only external testing will reveal how the model handles accents, overlapping speech in noisy environments, long contexts, factual reliability, and adversarial prompts. The lab has invited safety researchers precisely because a model that sounds convincingly human creates new risks of impersonation and social engineering. Those risks exist for any high-quality voice system; they are heightened when the model can be run offline and when its conversational style is difficult to distinguish from a person.

Second is the path from research preview to usable system. No weights, API, or pricing have been announced. A public waitlist exists, but the timeline for broader access is unclear. If the model remains available only to a small research group, its influence will be limited to the ideas it demonstrates. If weights or a local runtime appear, developers can begin to measure latency, memory use, and failure modes on real hardware.

Third is integration with broader agent systems. A speech model that stays on-device can serve as the interaction layer for tools that run locally or that call out selectively. It can also sit alongside the agentic systems whose memory, focus, and multi-step capabilities have been the subject of recent coverage of Meta’s Muse Wonder and Anthropic’s internal agent-to-human work ratios. The combination of natural voice and persistent agents raises both usability questions and new containment questions.

Finally, the efficiency claim deserves scrutiny. Training on one GB200 NVL72 is a concrete number. Whether the resulting model generalizes as well as text-trained systems of comparable inference cost is an empirical matter. If it does, the approach could influence how future voice and multimodal models are trained. If it does not, the demonstration still shows that full-duplex, speech-only training is feasible at a scale smaller than the largest language-model runs.

Conclusion

Dessa does not replace the large language models that currently power most AI products. It does not yet ship as a downloadable tool. What it does is make a specific architectural bet concrete: that a model can learn the structure of conversation from sound alone, that it can listen and speak at the same time, and that the resulting system can run on hardware a single person can carry. Those three properties together challenge the assumption that every voice interface must route through a remote text model.

The immediate next steps are external evaluation and clearer availability. Until then, the research preview functions as a proof that the information discarded by transcription is recoverable, and that the latency and privacy costs of always-on cloud voice are not inevitable. For anyone building interfaces that need to feel like conversation rather than like a sequence of transcribed queries, the existence of a speech-only, locally runnable full-duplex model is a data point that is hard to ignore.

Related TechMash reading includes analyses of open-weight progress with Kimi K3.1 Pro, local inference with Muse Spark Open on Apple silicon, agent capabilities in Muse Wonder, institutional responses to agents in banking systems, and AI-assisted security findings in PHP2Shell.