Comparison

Dessa AI vs NVIDIA PersonaPlex: Speech-Only vs Custom Personas

2026-11-04 · 7 min read · By DessaAI Daily

If Dessa asks whether a voice model should learn from speech alone, NVIDIA PersonaPlex asks a different question: how much control should you have over who the voice model pretends to be? Both are full-duplex systems that listen and speak at the same time. Almost everything else about them — training, openness and purpose — points in opposite directions.

Dessa is Block's speech-to-speech model, announced October 23, 2026 as a research preview with no weights, API or price announced. PersonaPlex is NVIDIA Research's full-duplex conversational model, built on Kyutai's Moshi architecture. This comparison uses Block's official Dessa facts and NVIDIA's own PersonaPlex research page, checked today.

How each system is built

Block says Dessa was trained from scratch on human conversation audio only — no text, no pretrained models. It learned language from hundreds of times fewer words than text large language models need, was trained on a single NVIDIA GB200 NVL72 system, and runs faster than real time locally on a MacBook Pro with an M4 Max chip or better.

NVIDIA describes PersonaPlex as built on the Moshi architecture, with the Helium text language model underneath providing semantic understanding. That is the opposite starting point: a proven text foundation first, then speech. NVIDIA lists PersonaPlex at 7 billion parameters — a figure Block pointedly has not published for Dessa at all. For the Moshi foundation both discussions share, see our earlier Dessa AI vs Moshi comparison.

The persona idea: a role you can prompt

NVIDIA's headline idea is control. PersonaPlex takes two prompts: a voice prompt — an audio embedding that captures vocal characteristics, speaking style and prosody — and a text prompt that defines a role. NVIDIA's examples range from a wise assistant to a customer service agent to a fantasy character, with the promise that the model maintains the chosen persona throughout a natural, full-duplex conversation.

Dessa makes no such promise, because Block is not presenting it as a character engine. In Block's live, unsteered recordings, Dessa chooses what to say, when and how, down to its own voice and personality. It laughs mid-sentence, backchannels, finishes a speaker's sentence, and occasionally gets things awkwardly wrong — in one recording it called a speaker “Ronaldo”. Where PersonaPlex gives a developer a dial for identity, Dessa presents an identity that emerged from training. Those are genuinely different product philosophies, not just different features.

Open weights vs a gated preview

Today, access is not close. NVIDIA describes PersonaPlex as a released research model built on Moshi, and third-party model records describe its code and weights as available under open licensing terms — check NVIDIA's own project page for the current licence before any commercial use, because terms matter and can differ for code and weights.

Dessa has released nothing comparable. Block has published no weights, no download, no API, no parameter count and no benchmarks, a gap we covered in Dessa AI open weights. Early access goes first to a small group of AI safety researchers, and everyone else joins the waitlist at block.xyz/dessa. Block, a payments company behind Square and Cash App, has chosen caution first; NVIDIA, a research and platform company, has chosen release and inspection. Neither choice proves which model sounds better.

Conversation feel and what can actually be tested

Both systems claim the core full-duplex behaviours: listening while speaking, knowing when to pause, interrupt or backchannel. NVIDIA says PersonaPlex learned not only the content of speech but the behaviour associated with it — when to pause, interrupt or say “uh-huh”. Block makes a strikingly similar claim for Dessa, with demos as its evidence.

The difference is verifiability. A researcher can download a Moshi-based system, run PersonaPlex-style experiments, measure latency and publish results others can repeat. A researcher outside Block cannot yet do any of that with Dessa. Block's demos are impressive and, by Block's account, live and unsteered — but they are still Block-selected recordings, with no independent evaluation published. Our Dessa AI explainer sets out what is claimed and what remains unproven.

Which matters for you?

If you are a developer or researcher who wants a full-duplex model you can obtain, inspect and shape into a role today, PersonaPlex and its Moshi lineage are the practical path. If you are watching for a breakthrough in how machines learn language itself, Dessa is the experiment to follow: a model that may show how much of conversation lives in sound alone, with no text foundation underneath it.

PersonaPlex extends a known architecture with a powerful new control. Dessa rejects that architecture's foundation entirely. The first approach can be tested now. The second, if it holds up, changes the question everyone else is asking.

Sources: Block's official Dessa page (block.xyz/dessa); NVIDIA Research's official PersonaPlex page (research.nvidia.com); Kyutai's Moshi materials; The Verge and TechCrunch, October 2026.

Frequently asked questions

← More Dessa AI articles Join the official waitlist ↗