Comparison

Dessa AI vs Gemini Live: Speech-Only Experiment vs Google's Live Voice

2026-10-31 · 7 min read · By DessaAI Daily

Google's Gemini Live is what a voice AI looks like when one of the largest technology companies ships it as a product: available now, connected to tools and services, and built on a foundation model that has read much of the written world. Block's Dessa, announced October 23, 2026, is almost the mirror image — a small lab's research preview that has read nothing at all, learned only by listening, and cannot be used by the public yet. Comparing them is less about crowning a winner than about seeing two opposite theories of how machines should learn to talk.

How each system is built

Gemini Live is Google's live conversational mode, built on the Gemini family of multimodal models. Those models are trained on text and other media at enormous scale, which is why a Gemini conversation can draw on broad factual knowledge, work with documents and images, and connect to tools and Google services. Voice, in this architecture, is a natural interface layered onto a very large, text-grounded intelligence.

Dessa deletes that foundation entirely. As we set out in what Dessa AI is, Block trained it from scratch on human conversation audio only — no text, no pretrained models. Block says 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. Where Gemini starts from reading and learns to speak, Dessa starts from listening and has never read a word.

Conversation feel: the ground Dessa chose to fight on

Block's whole case rests on conversational quality. Dessa is full-duplex: it listens while it speaks, so it can backchannel with "yeah!" and "got it!", laugh mid-sentence rather than after a pause, hear sarcasm and frustration, sigh, adjust its volume when asked — a behaviour Block says nobody explicitly taught it — and handle interruptions the way a person on a phone call does. We explain that mechanism in what full-duplex means.

Gemini Live also aims at natural, interruptible conversation, and unlike Dessa you can test it yourself today — that difference in testability matters. A claim you can check in five minutes on your own phone is worth more, evidentially, than a claim resting on a vendor's selected demo recordings. Block's demos are live and unsteered, but Block chose them, reviewers noted occasional awkwardness such as Dessa calling a speaker "Ronaldo", and no independent head-to-head evaluation of Dessa against Gemini Live exists.

Knowledge, tools and getting things done

Here the comparison is not close, and pretending otherwise would be dishonest. A text-and-multimodal foundation gives Gemini Live documented strengths Dessa has not claimed: factual breadth from vast reading, the ability to work with text and visual input, and tool use inside a large product ecosystem. If your goal today is a voice assistant that answers questions, summarises material and acts inside apps, Gemini Live is built for exactly that.

Dessa, by design, has never read a book, a manual or a webpage. Block has published no benchmarks, no parameter count and no tool-use or API announcement. Its knowledge is whatever spoken conversation taught it. That may prove surprisingly deep — people learn a great deal by listening — but it is unmeasured, and anyone who needs checkable, factual, tool-using work today should not plan around an unmeasured research preview.

Where the conversation happens: cloud versus local

Gemini Live is a cloud service: your audio is processed within Google's infrastructure under Google's policies, in exchange for scale, constant updates and access from ordinary devices. Block says Dessa's inference runs locally on high-end consumer hardware. If that form reaches users, conversation audio would not need to leave the device for the model to respond — a meaningful privacy and cost difference, and one we explored in why Dessa runs on a laptop. For now it remains a demonstrated capability, not a shipping product form: Block has released no download, no weights and no phone support announcement.

Price and availability

Gemini Live is available today through Google's apps and developer surfaces, under Google's published consumer and developer terms. Dessa has no price, no paid tier, no API and no public release date announced. The waitlist at block.xyz/dessa is free, and early access goes first to a small group of AI safety researchers — deliberately, because a voice this convincing raises impersonation and scam questions we covered in is Dessa AI safe? Anyone selling Dessa access today is not Block.

Which should you use?

Use Gemini Live if you want a voice AI now: knowledgeable, tool-connected, available on your devices today, and testable by you rather than described to you. Watch Dessa if you care about the scientific question underneath the product race — whether a model that learned only from speech can match or beat text-grounded systems at the thing conversation actually is: timing, tone, emotion and overlap. Google's entry proves how far the text-first road already goes. Dessa asks whether a completely different road leads somewhere that road cannot reach. The demos are intriguing; the verdict belongs to independent testing that has not happened yet.

Sources: Block's official Dessa page (block.xyz/dessa); Google's public Gemini Live materials as generally described; The Verge and TechCrunch coverage, October 2026.

Frequently asked questions

← More Dessa AI articles Join the official waitlist ↗