Updated October 25, 2026.
Dessa AI is a conversational speech model developed by a research team at Block, the company co-founded by Jack Dorsey. It is designed to listen and speak in real time, including while the other speaker is still talking. The team announced it on October 23, 2026 (UTC). Dessa team announcement
The interesting part is how it learned. Block says Dessa was trained entirely on human conversation audio, without text or pretrained models. That makes it an experiment in learning language through sound. Block's Dessa research preview
This guide explains that approach, what the demos show, how Dessa AI compares with ChatGPT Voice, and what you can actually access today.
1. Why do words alone miss part of a conversation?
Imagine a friend says, "Oh, great."
With a bright voice, it might mean good news. After a long sigh, the same words could mean the opposite. A transcript can preserve the sentence while leaving you to guess what the speaker meant.
Timing matters too. Someone who pauses may be searching for a name, inviting a response, or trying to finish a difficult thought. A useful voice assistant needs to decide whether to speak, wait, or offer a brief acknowledgment.
These details are often called prosody and conversational timing. Prosody includes pitch, stress, rhythm, and intonation. Conversational timing includes pauses, overlapping speech, and the exchange of turns.
Dessa's research premise is that an AI can learn these aspects together with language by training on spoken interaction. It treats the sound of a conversation as meaningful information.
2. How does Dessa AI work?
A conventional voice assistant can use three separate stages:
- Speech recognition converts the user's audio into text.
- A language model generates a written answer.
- Speech synthesis turns that answer into audio.
In this arrangement, the central language model may receive the words without directly receiving the speaker's tone or other sounds. OpenAI described this limitation when explaining the older ChatGPT Voice pipeline in its GPT-4o announcement. OpenAI's explanation of the earlier voice pipeline
Speech-to-speech describes a different interaction: audio goes into the model, and audio comes out. Dessa combines that interaction with Block's stated speech-only training approach.
Two ideas need to stay separate here.
Speech-to-speech describes the interaction. Speech-only describes the training material. A model can accept and generate audio while also learning from text or other modalities.
Block has not supplied a detailed architecture paper on the launch page. The diagram is therefore a conceptual explanation, not a map of Dessa's internal components.
3. What does full duplex mean?
Think of a phone call. You can hear the other person while you are speaking, and they can respond before you have completely finished.
That is the basic meaning of full duplex: listening and speaking can happen at the same time. The Dessa team identifies this as a core capability of the model. Dessa lead Ragavan Thurairatnam's announcement
For example, imagine saying, "The meeting is on Thursday—actually, Friday."
A conversational system should be able to incorporate the correction. During a longer explanation, it might also offer a short "mm-hm" without trying to take over. Researchers call these brief listener responses backchannels.
The difficult part is choosing the right moment. Constant interruptions are frustrating. Waiting through every pause can feel unresponsive. Full duplex enables overlapping interaction; good judgment about when to use it is a separate challenge.
4. What do the Dessa AI demos show?
Block's launch page includes conversations about a release announcement, remembering a movie, an anniversary surprise, and everyday topics. It highlights changes in loudness, laughter, sentence completion, and expressive delivery. Official Dessa recordings
Those examples help explain what the project is trying to improve: the rhythm of a conversation, including reactions that happen before a sentence ends.
They also leave room for awkwardness. In his coverage, The Verge's Jay Peters noted examples where Dessa used an unexpected name or wandered onto a movie tangent. Natural delivery and consistent conversational accuracy are different qualities. The Verge's coverage of Dessa
A model can sound surprised without having feelings. It can laugh at an appropriate moment while still misunderstanding a fact. Expressive audio should be assessed alongside the content of the answer.
5. Dessa AI vs. ChatGPT Voice: what is actually different?
It would be misleading to describe every version of ChatGPT Voice as a speech-recognition system attached to a text chatbot.
OpenAI introduced GPT-4o in 2024 as a model trained across text, vision, and audio. Its current GPT-Live documentation also describes simultaneous listening and speaking, with a backend handling delegated reasoning and tools. GPT-4o's multimodal design, GPT-Live documentation
A more useful comparison separates three approaches:
| Dimension | Dessa AI | Conventional voice pipeline | Other native voice systems |
|---|---|---|---|
| Main interaction | Speech to speech | Speech recognition, text response, speech synthesis | Can process and generate audio directly |
| Training approach | Human conversation audio only, according to Block | Depends on the models used at each stage | May combine speech with text or other modalities |
| Listening while speaking | Full duplex, according to the team | Depends on the surrounding implementation | Supported by systems such as GPT-Live and Moshi |
Full duplex also predates Dessa. Kyutai's 2024 Moshi paper describes a full-duplex speech framework built from a text language model backbone, with parallel speech streams and text-assisted generation. Moshi research paper
Dessa's distinguishing claim is its training entirely from human speech without a pretrained model. That is a specific research direction. It does not establish that Dessa outperforms ChatGPT at reasoning, factual questions, or completing tasks.
6. Can you use Dessa AI now?
As of October 25, 2026, the official site presents a research preview. Its public material does not provide a general chat experience, API documentation, or a downloadable model package. Dessa's official access information
The team is inviting AI safety researchers to explore the model, with the website describing access for a small group. The team's researcher invitation
| Item | Status on October 25, 2026 |
|---|---|
| Recorded demonstrations | Available on the official website |
| General public conversation access | Not available through the reviewed launch page |
| Public API | No documentation or public endpoint provided |
| Downloadable model weights | No public package provided |
| Pricing | Not published on the launch page |
| Safety researcher access | Limited early access by application |
Can Dessa AI run locally on a Mac?
Jack Dorsey's stated hardware guidance is a MacBook Pro with M4 Max or better. This describes a claimed local inference capability; it does not give readers a model they can install today. Dorsey's hardware guidance
The launch page does not specify minimum memory, a complete supported-device list, or installation instructions.
For access updates, start with the official Dessa website. Follow the links Block provides when a public release becomes available.
7. Why does Dessa AI matter?
The research question extends beyond whether a voice sounds convincing.
If language and conversational behavior can be learned effectively from speech alone, researchers gain another way to study how training material shapes an AI's abilities. They can examine what the model learns about meaning, timing, factual knowledge, and interaction without starting from a text-trained system.
Possible applications include language practice, conversational companions, and voice agents that handle interruptions more naturally. These are potential uses of the approach, rather than announced Dessa products.
Each would need its own evaluation. A language tutor needs accurate corrections. A customer service agent needs reliable access to company information and tools. A companion needs appropriate boundaries. Fluent delivery cannot establish those capabilities by itself.
Useful future tests would measure interruption handling, performance across accents and noisy rooms, factual reliability, and the stability of longer conversations.
8. Frequently asked questions
Who created Dessa AI?
A small research lab at Block developed Dessa. The team is led by Ragavan Thurairatnam and includes former members of the Toronto deep learning startup also called Dessa. Block's team description
Is Dessa AI open source?
The reviewed launch material does not provide public weights or a model license. A research announcement alone does not establish an open-source release.
Is there a Dessa AI API?
No public API documentation or endpoint is provided in the official launch material as of October 25, 2026.
Is Dessa AI better than ChatGPT Voice?
The available demonstrations do not establish an overall winner. A fair comparison would test both systems under the same conditions and measure conversational timing, answer quality, and task completion separately.
Dessa AI offers an early look at learning conversation directly from human speech. For now, readers can explore the official recordings and follow the research. Broader access and independent evaluations will make it possible to judge how well the approach works beyond the selected demos.
Top comments (0)