Voice in. Voice out. No text in between.

On 23 Oct 2026, Block (the company behind Square and Cash App, run by Twitter co-founder Jack Dorsey) showed Dessa: a speech model trained from scratch on nothing but recorded human conversation. No text. No pretrained model underneath.

Speech-to-speech Full-duplex Trained only on audio Faster than real time Runs on a MacBook Pro M4 Max Waitlist only

block.xyz/dessa, the official launch page.

Hears words

your voice → text → LLM → text → voice

Tone, laughs, sighs and sarcasm get thrown away at the “text” step.

Hears you

your voice → Dessa → voice

Audio in, audio out. Emotion and timing survive the whole way.

It listens while it talks.

“Full-duplex” means Dessa hears you and speaks at the same time, like a person on a call. Block’s demo marks the moments a normal voice bot can’t do:

Dessa talking with another AI (GPT-Live-1). The labels are Block’s: interruption, patience, fast backchanneling.

Laughs mid-sentence

When something is funny, it laughs right then, not politely after you finish.

Backchannels

“yeah!”, “got it”, “ooh” while you speak, so you know it’s with you.

Knows when to cut in

It learned turn-taking from real conversations: when to wait, when to jump in, when to finish your sentence.

Hears emotion

Sarcasm, a frustrated sigh, a voice cracking. Ask it to speak up and it gets louder.

Try this out loud: “I didn’t say she stole my code.” Stress a different word each time. Same words, different meaning. Text can’t carry that. Speech can. That’s the whole bet behind Dessa.

4 steps. 10 minutes.

Listen to the live recordings

Open block.xyz/dessa and press play. Every clip is a real person talking to Dessa, nobody steering it. Start with “Remembering a movie” (it finishes a title the person forgot), “An anniversary surprise” (it gasps “oh my god”) and the one where it talks to ChatGPT’s voice model.

Join the waitlist

Same page, “Join the waitlist” at the top. It’s a research preview, so access is invite-only for now. The download button says “coming soon”.

Researcher? Ask for early access

Block is giving a small group of AI safety researchers early access. Scroll to “Calling all safety researchers!” and apply with your university, company or lab email (a Google Scholar or blog link helps).

Follow for new clips

The team says more recordings will be shared on X. Follow the Dessa lead, Ragavan Thurairatnam (@RagavanThurair).

From Block’s page: audio in, audio out, full-duplex, runs locally on a MacBook Pro M4 Max or better.

Not out yet: no public API, no pricing, no model weights, no benchmarks. Any “Dessa app” or “Dessa API key” you see selling today is not from Block.

Voice is the next interface. Start before Dessa ships.

You can’t call Dessa from code yet. But voice agents already work well enough to ship real products today, and when emotional models like Dessa open up, you swap the voice and keep the product.

Works today

Phone calls that get answers

Confirm appointments, chase leads, ask a shop if something is in stock. The AI calls, talks, reports back.

Works today

Voice support line

Answer the same 20 customer questions at 2 AM, in the customer’s language.

Works today

Speaking practice

Interview prep or language practice with a partner who never gets bored.

Where emotion matters

Care & check-ins

Elder check-in calls, coaching, mental-health triage: this is where “hears how you feel” changes everything.

My take: I used to tell people AI won’t turn on us because it has no emotions. Now we’re deliberately putting emotion in. That’s exciting and a little scary. Block says the same: safety is one of their three reasons for building it.

• Pick one call people hate making
• Write the script in plain words
• Test it on 5 real calls
• Read every transcript
• Fix the opening line first
• Charge per minute, not per month

Next →

Want an AI that already makes real phone calls? Turn the page.