Disclosure: this article contains affiliate links to ElevenLabs and Murf AI. If you sign up through one, I may earn a commission at no extra cost to you. It doesn't change the verdicts — and the main subject of this piece, Dessa, isn't something anyone can sell you yet.
You can't try Dessa. I want to get that out of the way in the first paragraph, because most of the coverage I've read buries it.
Block announced Dessa on October 23. It's a full-duplex, speech-to-speech model from Jack Dorsey's company, and the demos genuinely sound like a person — pauses, laughter, the little "yeah, got it" noises we make without thinking. But there is no API. No pricing. No weights, no download, no benchmark sheet. Access right now is a public waitlist, plus early access for a small group of AI safety researchers. Everything anyone outside Block "knows" about how good it is comes from Block's own demo recordings.
So why write about a product you can't use? Because Dessa is the clearest signal yet that the way voice AI is built is about to change — and if you ship voice features, support calls, or audio content, the change lands on your desk whether you adopt it or not.
The pipeline you're using now (and what it throws away)
Almost every voice AI you've talked to is three separate systems bolted together:
- Speech-to-text transcribes what you said.
- A language model reads the transcript and writes a reply.
- Text-to-speech reads the reply out loud.
It works. It also deletes most of the conversation. Your sigh, your hesitation, the sarcasm in your voice, you talking over the bot because it's heading the wrong way — none of that survives transcription. Text carries words. It doesn't carry how they were said.
Block's argument, in its announcement, is blunt: the most powerful AIs in the world can't hear how you feel, and that makes voice interaction both worse and, in their framing, potentially dangerous. You can agree or not, but the diagnosis is fair. Anyone who has shouted "representative" at a phone tree knows the feeling.
What Dessa actually is
Per Block's announcement and the launch coverage (TechCrunch, Oct 24):
- Trained from scratch on human conversation only. No text. No pretrained language model underneath. Everything it knows, it learned from recorded speech — hundreds of times fewer words than a text LLM trains on, by Block's account.
- Full-duplex. It listens and speaks at the same time. You can interrupt it mid-sentence, laugh, backchannel — and it reacts the way a person would, because those behaviours live in the training data.
- Emotion is in the model, not in markup. Tools like ElevenLabs v4 or Fish Audio steer feeling through voice tags. Dessa's emotion is baked into how it speaks, according to Block — it doesn't need to be told to sound tired; in one demo, told it was tired from work, it sounded tired.
- Trained on a single NVIDIA GB200 NVL72 rack, and Block says inference runs faster than real time locally on a MacBook Pro with an M4 Max or better. No cloud round-trip required.
- Knowledge comes from conversation, not retrieval. Most speech models pair with an LLM for knowledge. Dessa's knowledge stems from its conversational training data instead. Hold onto that detail — it matters below.
The lab lead, Ragavan Thurairatnam, called it a "GPT-3 moment for speech." That's a big claim from an interested party, and right now it's unfalsifiable either way: no parameter count, no published evaluations, no independent testing. The Verge listened to the demos and found them mostly natural but occasionally awkward — in one, the model mistook the speaker for someone named "Ronaldo." Plan accordingly.
Full-duplex vs. what you can actually buy
| Dessa (Block) | GPT-Live (OpenAI) | Typical voice-agent stack | |
|---|---|---|---|
| Architecture | Single speech-to-speech model | Full-duplex model | STT + LLM + TTS pipeline |
| Text in the middle? | No — audio in, audio out | No | Yes — everything passes through a transcript |
| Interruption / overlap | Native, per Block | Native | Bolted on via VAD and barge-in handling |
| Knowledge source | Its conversation training | A frontier model lineage | Whatever LLM you plug in — swappable |
| Availability | Research preview, waitlist | Shipping product | Build it today from APIs |
| Benchmarks | None published | Vendor-reported | You measure it yourself, in production |
The honest read: full-duplex isn't a Block invention — OpenAI's GPT-Live got there as a product first. Dessa's bet is different. It's that removing text entirely produces conversation that feels more human than any pipeline, however good the components get.
What this changes if it works
Call centres. This is where full-duplex bites first. Real callers interrupt, talk over prompts, cry, get sarcastic. A model that hears all of that — and doesn't need a neat transcript to respond — handles the messy 20% of calls that current bots fumble into "let me transfer you." But here's the catch nobody's demo answers: Dessa's knowledge comes from conversation, not from your help centre, your CRM, or last Tuesday's pricing change. A voice that sounds human but can't look anything up is a very charming dead end. The winning support product is probably full-duplex ears and mouth on top of a retrievable brain — which is, ironically, a pipeline again, just a better one.
Creators. Scripted text-to-speech is fine for narration. It is bad at dialogue — two voices reacting to each other, timing jokes, leaving space. A conversational model that generates natural back-and-forth could do podcast banter, character scenes, and interview-style content that today needs two humans or heavy editing. We're not there yet (no product!), but this is the use case I'd watch.
App builders. Voice stops being a chat window with a microphone button. If the model can listen while it speaks, your interface design changes: no push-to-talk, no waiting for the beep, no "sorry, I didn't catch that" loops. The design constraint moves from transcription accuracy to conversation management — turn-taking, interruption policy, when the agent should just shut up and listen.
The catches, collected in one place
- No benchmarks. Zero. Every capability claim traces back to Block's demos.
- Knowledge ceiling. Conversation-trained knowledge is frozen at training time and shallow by design. Fine for chat, dangerous for facts.
- The manipulation problem is real. A model that hears emotion can also perform it — synthetic warmth, synthetic urgency. Block giving safety researchers first access suggests they know this. Anyone deploying voices like this should assume disclosure rules (and yes, ad laws — see what's already happening in Pennsylvania) will follow the capability.
- Demos are not products. A waitlist, a rack-sized training run, and a handful of recordings are a research announcement, not a release.
What you can use today
Until Dessa (or something like it) ships, voice work still runs on pipelines. Two tools cover most of the actual jobs, and I'll be specific about which job each one is for — and what each can't do.
ElevenLabs — the default pick for both studio-quality text-to-speech and real-time conversational agents. The voices are the strongest I've found for narration and character work, and their Agents platform handles the STT/LLM/TTS plumbing for phone and web voice agents. The catch: it is the pipeline architecture — emotion is steered through controls and tags rather than learned, turn-taking still needs tuning, and per-minute agent costs add up fast at call-centre volume. Check current pricing before you commit; it changes often. Try ElevenLabs
Murf AI — a studio voiceover tool, and I mean that as a boundary, not a compliment or a criticism. If your job is e-learning narration, explainer videos, or ads read in a consistent voice, Murf's editor is faster and simpler than ElevenLabs. The catch: it is not a conversational agent. Nothing you build in Murf will listen to anyone. Buy it for narration, never as a step toward a Dessa-style product. Try Murf AI
And if you just want to feel full-duplex today, GPT-Live inside ChatGPT is the shipping example — no purchase needed, ten minutes of conversation will teach you more about the interaction model than ten articles (this one included).
My take
Dessa is a direction, not a product. Treat it that way. If you're shipping voice this quarter, the boring pipeline tools above are your answer, and your money is better spent on turn-taking tuning and good fallback design than on waiting for a waitlist. If you're building for 2027, design your voice UX around interruption and overlap now — because the first widely available speech-native model, whoever ships it, will make push-to-talk interfaces feel like dial-up.
For the folks already running voice agents in production: is awkward turn-taking actually costing you users — or is it the least of your problems compared to wrong answers and latency? I genuinely want to know where the pain ranks.
Sources: Block's Dessa announcement (Oct 23, 2026); TechCrunch, "Block's new voice model sounds eerily human" (Oct 24, 2026); The Verge's demo impressions as reported in launch coverage. All Dessa performance and hardware claims are Block's own and are labelled as such — none have been independently verified, and I have not used the model, because nobody outside Block's early-access circle publicly has.
Top comments (0)