Block's Dessa voice AI is a full-duplex speech-to-speech model: it listens while it talks, so callers can interrupt it, and it can respond with pauses and backchannels instead of waiting for silence and reading out a reply. For businesses, that points to voice agents that sound less like a phone menu and more like a person on the line.

Dessa is not publicly available: public waitlist only, early access limited to a small group of AI safety researchers, and no API, price, weights or benchmarks published by Block. This is not a buying guide. It explains what full-duplex changes, what it does not change, and how to prepare now using voice agents you can deploy today.

Block's Dessa Voice AI: What Full-Duplex Speech Means for Your Business

What Dessa actually is

Block's small in-house AI research lab unveiled Dessa on 23 October 2026. The lab is led by Ragavan Thurairatnam, reports to Jack Dorsey, and was founded by former members of the Toronto deep-learning startup Dessa.

Dessa takes audio in and produces audio out, with no transcription step and, per Block, no text model or large language model underneath. Per Block, it was trained from scratch on human conversation only — no text, no pretrained models — on a single NVIDIA GB200 NVL72 rack, and it runs locally on a MacBook Pro M4 Max or better, faster than real time. Those are Block's demo claims, not verified benchmarks, since Block has published none.

How most voice agents work today

A typical voice agent transcribes speech to text, generates a text reply, then reads it aloud. Each stage adds delay and loses tone, hesitation, laughter and emphasis. Dessa, as Block describes it, removes those middle steps.

Why full-duplex changes a phone call

A normal phone call is full-duplex: both people can speak at once, interrupt each other, and recover gracefully. Most automated phone systems are turn-based instead: you speak, you stop, the system thinks, the system speaks, and if you interrupt, it ignores you or starts again.

Dessa, as demonstrated, listens while it talks. That enables three behaviours turn-based systems struggle with:

• Interruption. A caller can cut in — "no, Thursday, not Tuesday" — without waiting for the agent to finish.
• Backchannelling. The short sounds people make while listening — "mm-hm", "right", "go on" — that signal attention.
• Natural pacing. Pauses, hesitations and emphasis that carry meaning, rather than a flat, evenly paced readout.

This is not cosmetic. If a caller cannot correct a date mid-sentence, they repeat themselves, get the wrong booking, or hang up and call a competitor.

What text pipelines lose — and why it costs you calls

The comparison matters because it shows up in caller behaviour, not just in technology:

• Transcribe-then-reply pipelines are proven, deployable and easy to connect to your systems, but they add latency, flatten emotion, and handle interruptions poorly. They are the right choice for most businesses today, precisely because they exist.
• Full-duplex speech-to-speech models like Dessa promise lower latency and more natural conversation, but they are not yet a product you can buy. Their behaviour with accents, noise, crosstalk and angry callers is unproven in public.

Do not wait for perfection. If you lose callers because nobody answers or simple bookings take too many steps, a well-built voice agent solves that today. A caller sounding rushed or frustrated is also data — a text pipeline usually discards it at transcription, while a speech-to-speech model hears it directly. Treat that as promising, not proven, until Dessa is tested live.

What Dessa does not change

A voice model, however natural, is only the conversation layer. Everything around it still has to be built:

• Booking systems. An agent that cannot check a real calendar or confirm a table is a demo, not a service.
• Accurate business data. Prices, hours and availability must come from your systems, not a model's memory. This is where AI automation and proper data connections earn their keep.
• Guardrails. What the agent may promise, what it must never say, when it hands over to a human, and how it handles complaints all need explicit rules.
• Disclosure and consent. Callers should know they are speaking to an AI. A more human-sounding voice makes disclosure more important, not less.

Where full-duplex would matter most

Appointment-heavy businesses benefit most: clinics, salons, restaurants and home services field calls where callers interrupt, correct details, and ask follow-up questions mid-sentence. Our guide to AI chatbots for clinics and healthcare covers similar ground on the text side.

Restaurants are another strong case: a group booking changes party size, dietary options and requests mid-call. A restaurant ERP connected to a voice agent turns that into a confirmed booking instead of a missed call. The same applies to logistics, where addresses and delivery windows need correcting in real time — see our guide to AI automation for logistics companies.

How to prepare now, without waiting for Dessa

Build the foundations any voice agent will need, text-based today or full-duplex tomorrow.

1. Audit your calls before you automate them

List the top ten reasons people call, grouped into bookings, questions, complaints and handoffs. Rare or judgement-heavy calls stay with humans; frequent, repeatable calls are automation candidates. Our guide on how to build an AI voice agent walks through this scoping step.

2. Fix the data first

Clean calendars, up-to-date service lists, and a single source of truth for prices and availability will improve any agent you deploy. This is where AI agents development overlaps with voice: the agent needs tools and permissions, not just a voice.

3. Deploy a voice agent you can measure

Start with your highest-volume, simplest call types, and measure answer rate, booking completion, handoff rate and complaints. For budgeting, see our AI calling agent cost guide for 2026 — treat figures there as planning guides, not quotes.

4. Keep chat and voice on the same brain

Many customers start on WhatsApp and finish on the phone, or vice versa. A WhatsApp bot and a voice agent sharing the same booking data give a consistent experience. See AI chatbot vs AI agent: a chatbot answers questions, an agent completes tasks, and voice callers expect task completion.

Should your business wait for Dessa?

For almost every business, no. Dessa is a research demonstration with a waitlist, not a product with a launch date, pricing, or integration path. Waiting means leaving missed calls unanswered. The businesses that benefit most when full-duplex models arrive will be those with clean data, working integrations, and measured voice deployments already in place.

Building on chatbot development and custom software development foundations you control means you are not locked into one model. When a full-duplex model ships with a real API, you adopt it as an upgrade, not a rebuild. That holds across markets, which is why we work across the United Kingdom, the USA, the UAE, Saudi Arabia and Pakistan.

For grounding agents in your own data, see our RAG chatbots business guide. Announcement coverage is in TechCrunch's report on Block's new voice model and the Indian Express report on Jack Dorsey's Dessa voice AI. Both describe demos, not a public release — Dessa is direction of travel, not a purchase decision.

Frequently asked questions

Can my business buy or deploy Block's Dessa today?

No. Public waitlist only, early access limited to a small group of AI safety researchers, and no API, price, weights or benchmarks published by Block.

What does full-duplex actually mean in plain English?

The system listens while it talks, like a person on a phone call. Callers can interrupt and correct themselves, instead of waiting for silence and speaking in turns.

Is Dessa a large language model with a voice added?

Per Block, no. Dessa is speech-to-speech — audio in, audio out, no transcription step, no text model or LLM underneath — trained, per Block, from scratch on human conversation only.

What should I build now if I want full-duplex voice later?

Build what does not change: clean booking data, working integrations, clear guardrails, disclosure and consent processes, and a measured deployment on your simplest, highest-volume calls.

Will full-duplex voice replace human phone staff?

Not for the calls that matter most. It can handle frequent, repeatable calls — bookings, hours, order status — and free humans for complaints and complex cases, with handoff rules defined in advance.