Dessa: the most authentic full-duplex speech model ever built

Block unveiled Dessa, an authentic full-duplex speech model trained entirely on raw human audio without text tokens or pretrained foundations, according to block.xyz's report. Developed inside a specialized research lab at Block led by Ragavan Thurairatnam, the system processes audio inputs directly into audio outputs in real time. Rather than relying on automatic speech recognition coupled with a text transformer and voice synthesizer, the architecture acquires conversational patterns, emotional inflection, and factual knowledge strictly by listening to recorded human speech.

This report covers the technical design of Block's audio-native model, its hardware requirements, and the operational differences between raw acoustic learning and traditional text pipelines. It is intended for software engineers, machine learning researchers, and product teams tracking conversational voice interfaces, edge inference, and multimodal interaction architectures. Readers will find details on training footprint constraints, edge execution targets on Apple Silicon, and the acoustic mechanics separating full-duplex systems from standard turn-taking conversational engines.

The most authentic full-duplex speech model: what actually changed

Conventional voice systems translate user speech into text, pass those tokens through large language models, and synthesize speech from the output text. That layered process strips out vocal acoustic signals like breathing, inflection, hesitation, and emotional pitch, yielding robotic cadence and perceptible latency. Dessa replaces that three-tier pipeline with a unified speech-to-speech architecture trained without text or pretrained weights. The model was trained from scratch exclusively on audio recordings of human conversations, learning vocabulary, conversational cadence, and contextual awareness directly from raw sound waves.

This architecture changes how artificial intelligence handles real-time conversational timing. Instead of waiting for a speaker to finish a sentence before generating a response, Dessa operates in full duplex, actively listening while simultaneously speaking. In recorded demonstrations, the model cut in with appropriate backchannel cues, laughed mid-sentence when a conversationalist made a joke, and raised its speaking volume immediately when asked to speak louder without separate programmatic prompting. By working directly on acoustic data, the model interprets tonal shifts that text tokenizers discard entirely.

The most authentic full-duplex speech model: how it works

Dessa operates as an end-to-end audio-in, audio-out system that learns language representations from acoustic variations rather than discrete text tokens. While text characters remain identical regardless of who types them, spoken words never repeat identically, requiring significantly more data bandwidth per word and demanding that the model resolve accents, pauses, and cadence variations directly. By capturing subtle stress changes across phrases where shifting emphasis alters sentence meaning, the system internalized syntax, humor, and conversational turn-taking from raw sound files rather than tokenized syntactic trees.

On the infrastructure side, Block trained Dessa from scratch using only a single Nvidia GB200 NVL72 rack, avoiding massive supercomputing clusters typically required for frontier language foundations. Because speech-only training bypasses extensive text corpuses containing programming languages and documentation, Dessa learned its vocabulary from hundreds of times fewer words than traditional text models. For production inference, the model runs faster than real time locally on an Apple MacBook Pro equipped with an M4 Max chip or better, eliminating server network latency during live dialogue.

The most authentic full-duplex speech model: why it matters now

Current automated conversational assistants frequently produce stilted user experiences because sequential translation between audio and text creates turn-taking lag. When acoustic data gets flattened into written transcripts, models become functionally deaf to the distress, urgency, frustration, or sadness in a speaker's voice. The engineering team behind Dessa highlighted this lack of emotional perception as a safety problem, noting that an assistant unable to detect vocal pain or panic might continue standard output scripts inappropriately when a human user requires immediate crisis intervention.

Beyond real-time user experience improvements, Dessa establishes a new research category for deep learning efficiency by demonstrating that factual retrieval and linguistic comprehension can emerge strictly from auditory perception. The model proved capable of completing forgotten movie titles and offering empathetic life advice using only facts acquired from audio data. This acoustic-first approach challenges the assumption that broad multi-billion-token text corpora are strictly necessary to build responsive conversational models, demonstrating that acoustic pretraining can teach models speech timing, expressive inflection, and factual knowledge concurrently.

The most authentic full-duplex speech model: who is affected

The immediate audience impacted by this release consists of artificial intelligence safety specialists, deep learning researchers, and conversational software developers. The research team, led by Ragavan Thurairatnam alongside researchers Louay Hazami, Hashiam Kadhim, Qi Alex Liu, Rayhane Mama, Aydin Polat, and Marc Tyndel, reports directly to Block cofounder, Chairman, and Block Head Jack Dorsey. With former startup personnel who developed generative voice systems since 2012, this unit operates outside standard enterprise voice practices, prompting engineering teams building customer contact platforms and personal assistants to reconsider cascading pipeline architectures.

Developers of edge applications and local desktop utilities also face direct implications as real-time audio inference migrates away from centralized cloud infrastructure. Proving that an end-to-end full-duplex speech engine can execute on an M4 Max MacBook Pro shows that low-latency voice intelligence can run entirely on client machines without sending private audio to remote server farms. Furthermore, independent model developers and third-party voice providers must now benchmark against systems capable of spontaneous interruptions, genuine conversational gasps, and dynamic acoustic adaptation during interactive sessions.

The most authentic full-duplex speech model: what to watch

Industry observers should monitor the outcomes of Block's initial safety evaluation program as external specialists probe the model for unpredicted risks. The Dessa team opened a waitlist for researchers, offering early access to evaluate safety vulnerabilities unique to models trained entirely without text guardrails or standard post-training reinforcement filters. Because raw sound training represents uncharted territory in deep learning, safety researchers will study how audio-native systems handle prompt manipulation, emotional entanglement, and spontaneous dialogue escalation during extended unscripted conversational sessions across diverse acoustic environments.

Future updates will reveal how effectively audio-only models scale linguistic transfer, learn secondary languages, and perform in-context reasoning compared to massive text architectures. Observers should track additional audio demonstrations planned for release on X, including further side-by-side conversational tests similar to Dessa's dialogue session with GPT-Live-1. As Block releases more performance benchmarks and technical documentation, the broader machine learning community will evaluate whether training audio models on single-rack hardware like the Nvidia GB200 NVL72 can be replicated for specialized industry voice applications.