Dessa is a voice model from Jack Dorsey trained only on human conversation instead of text, which is why it handles sarcasm, pauses, breathing, accents and comedic timing better than most text to speech. This post covers how the Dessa voice model by Jack Dorsey works, what it is for, and where to hear it.
Human conversation audio (no text)
|
v
Dessa voice model
|
+--> pauses, breath, laughter, timing
+--> sarcasm and emotional delivery
+--> accents kept, not flattened
|
v
voice agents | cold outreach | support callsKevin covered the release on his channel, so if you prefer to hear it explained out loud, watch the walkthroughs on the KevBuildsApps YouTube channel. The primary source is the project page at block.xyz/dessa, and the video breakdown lives at this link.
What is the Dessa voice model by Jack Dorsey?
Dessa is a conversational voice model: it produces speech that sounds like a person talking, not a person reading.
That distinction matters more than any benchmark number. Most text to speech systems start from written text and learn to map characters to sound. They get very good at pronouncing words. They get less good at the parts of speech that are not words: the half second before a difficult sentence, the small intake of breath, the flat tone that signals irony, the speed change when someone is excited.
Dessa takes the other path. It is trained on recordings of people talking to each other. That is the whole premise, and it is why the output lands differently when you listen to it next to a conventional model.
Jack Dorsey is the person behind it, and Block publishes the project at the link above. The work sits in the same family as the other things he has shipped: tools for talking to people, built around the way people actually communicate rather than around a clean technical abstraction of it.
For a quick sense of how fast this space is moving, our older post on Voxtral 3 and voice replies shows the same shift on the output side, where voice stops being a narration layer and starts being a channel you answer customers on.
How does Dessa work without any text training data?
Training on conversation audio alone means the model learns delivery as a first class signal, not an add-on.
Think about what a transcript throws away. A transcript of two people arguing is a flat wall of words. The audio contains who spoke over whom, where the argument softened, where someone laughed while pretending not to. If you train on transcripts, none of that reaches the model. If you train on the audio, all of it does, because the timing is in the waveform alongside the words.
The practical consequences you can hear:
- Sarcasm needs a specific pitch contour and a slight drag on the stressed word. Learning that from text is close to impossible, because text does not mark it.
- Pauses fall where a thinking person puts them, not where punctuation says. A comma and a real hesitation are different lengths.
- Breathing is present because the training audio has breath in it. The model does not know it is generating breath; it is reproducing the statistics of the recording.
- Accents stay. If the source speakers had a regional accent, the model reproduces the accent rather than smoothing it into a neutral broadcast voice.
- Timing carries jokes. Comedy is mostly rhythm, and rhythm only exists in the audio.
This is also why the model is worth watching for anyone building a voice agent. A cold call that opens with a flat synthetic greeting is dead in two seconds. A call that opens with a small natural breath and a normal pause is a different conversation entirely.
Why Dessa sounds real on sarcasm, pauses and breathing
The realism is a side effect of the training data, not a post processing filter.
Most “make the voice sound human” tricks are applied after generation: insert a pause, add a breath sample, pitch shift a syllable. Listeners notice because the pieces do not fit. The pause arrives in the wrong place and the breath is identical every time.
When the model has only ever heard whole conversations, it produces whole delivery. The breath is where a breath would be for that sentence. The pause lands where the speaker was working something out. That coherence is what your ear reads as a person.
There is a tradeoff, and it is worth saying plainly. A model that produces natural hesitation will sometimes hesitate when you wanted a crisp line. Natural delivery is not the same as controlled delivery, and if your use case is a 15 second ad read that must fit an exact slot, a more controllable model may serve you better.
How does Dessa compare to text to speech models like Voxtral 3?
They aim at different jobs, and the right pick depends on whether you need performance or delivery.
| Question | Conversational model (Dessa) | Text to speech model (Voxtral 3 and similar) |
|---|---|---|
| Input | Conversation context and intent | Written text, sometimes with tone tags |
| Training signal | Human conversation audio | Text paired with speech, usually read speech |
| Strength | Delivery, timing, emotion, accents | Precision, repeatability, script control |
| Best fit | Voice agents, real time replies, outreach calls | Narration, explainers, scheduled content |
| Hard part | Scripted exactness | Naturalness on emotional lines |
If your goal is a reply to an Instagram or Facebook comment that sounds like you, the Voxtral style route is the practical one today, and our guide to agent replies in your own voice walks through the workflow. If your goal is a phone call with a stranger who should not immediately know they are talking to a model, Dessa is the more interesting direction.
What can you build with Dessa: voice agents and cold outreach
The obvious use cases are conversations where the other person is listening for whether you are real.
Three that fit naturally:
- Inbound voice agent. Someone calls about pricing. The agent greets, listens, hesitates like a person considering the question, and answers. The hesitation is the product.
- Cold outreach. A first call where tone decides whether the call continues. A natural opener buys you the next ten seconds.
- Follow up after a social conversation. A customer comments on your post, then gets a real call. Here the voice layer sits on top of an existing pipeline.
That third one is where Creator OS fits, and the fit is narrow enough to state honestly. Creator OS is where the conversation starts: it runs your posts, comments and DMs across 11 platforms plus Skool, and it ships a hosted MCP server and a CLI so an AI agent can run those accounts. A comment to DM funnel pulls a commenter into a private thread with one command:
creatoros automations:create \
--name "guide funnel" \
--accountId acc_123 \
--postId post_456 \
--keywords "guide,link,send it" \
--matchMode contains \
--dmMessage "Here is the guide you asked for. Can I call you this week?" \
--commentReply "Sent you a DM."The DM is where Dessa style voice would take over: the customer replies with a number, and the voice agent makes the call. Creator OS handles the public conversation and the inbox; the voice layer handles the phone. Our Instagram to Claude sales engine post covers the same funnel shape from a different angle.
If you build agents rather than pipelines, the MCP server post explains how the connection works in Claude, ChatGPT, Cursor and the rest.
How do you get early access to Dessa and test it?
Go to the source first: the project page at block.xyz/dessa, and the video breakdown at this link.
Early access details can change quickly, so treat the Block page as the only authority on availability, and do not rely on secondhand summaries. What you can do today is prepare, and the preparation is the same whether access opens next week or next quarter:
- Write the script you actually want. Not a marketing paragraph. Two or three sentences of a real opening call, with the awkward parts left in.
- Record the human version. Say it yourself, badly, on a phone. That recording is your reference, and it is what you compare against.
- Test on the hard line. Every script has one line that a synthetic voice ruins. Find it first. Sarcasm, a name, a number.
- Keep a control. Run the same script through a text to speech model you already use, so you are comparing rather than admiring.
For the agent side of the build, the open source Social Agents project (github.com/kevinbadi/social-agents) is a working harness that posts, replies to comments and DMs, and reports, with your own model. It is a reasonable place to slot a voice step into an existing loop.
What are the risks and limits of a conversation-only voice model?
A model with no text training data is excellent at conversation and weak at control, and both halves deserve attention.
The limits worth planning around:
- Exactness. Product names, prices and phone numbers need to be right every time. Natural delivery can drift on unfamiliar words.
- Consistency. If you need the same line identical across a thousand calls, a conversational model fights you.
- Script changes. Naturalness learned from conversation does not automatically transfer to a script nobody would ever say out loud.
- Disclosure. Rules on AI voice in calls differ by country and platform. Check them before you dial, and say when a caller is talking to a system if the rules require it.
- Consent and identity. A model this good at mimicry raises real questions about whose voice went into the training audio and who can clone whom.
The safe pattern is to keep the voice layer narrow. Let it handle the opening and the emotional beats. Hand off to a human the moment the call turns commercial, and keep a written record of what was said. That is a boring answer, and it is the one that survives contact with a legal team.
FAQ
What is the Dessa voice model by Jack Dorsey?
Dessa is a voice model from Jack Dorsey that is trained on human conversations rather than on text. That training is what lets it handle sarcasm, pauses, breathing, accents and comedic timing. Block publishes the project at block.xyz/dessa.
Is Dessa trained on text or only human speech?
It is trained on human conversation audio, with no text training data involved. Because the timing and breath are in the audio, the model picks them up alongside the words. A transcript would have thrown exactly that information away.
Can I use Dessa for cold calling and voice agents?
That is the natural fit, because both depend on the listener believing they are talking to a person. Natural delivery buys you the first few seconds of a cold call. Prepare a short real script and test the one line that synthetic voices usually ruin before you build anything larger.
How does Dessa fit next to text to speech tools?
Text to speech is built for turning a written script into speech, and Dessa is built to sound like a live conversation. That makes Dessa worth trying where delivery and emotion matter, and a script driven model worth keeping where exact wording matters. Decide by asking which one your use case needs more.
Where can I try Dessa from Block?
Check block.xyz/dessa for the current status, since early access details can change. The video breakdown is at this link, and Kevin’s channel covers the walkthrough: youtube.com/@KevBuildsApps.
Get started with Creator OS
Build the conversation side first: comments, DMs, and the funnel that turns a commenter into a caller. Creator OS runs that across 11 platforms plus Skool, and the API, MCP server and CLI are included in every plan.
Create an account at creatoros.ca/sign-up. Plans start at $14.99/month for 11 platforms plus Skool.
Then read the docs at creatoros.ca/docs for the setup, the post options and the automation calls. Watch the build videos on KevBuildsApps, and see the platform in motion in the launch video.
Get started for $0
Free trial for new subscribers. Cancel anytime.