Block’s Dessa Voice Model Brings Humanlike Emotion to Local AI Conversations

Table of Contents

• You might want to know
• Block’s Dessa Voice Model
• Key Insights Table
• Afterwards...

You might want to know

• How does Block’s Dessa model aim to make conversations with voice AI feel more natural?
• What questions about safety and access remain as the model is developed?

Block’s Dessa Voice Model

Block, the company led by Jack Dorsey, has introduced a voice model from its research lab called Dessa. The model is designed to run locally and, according to the company, can produce speech that sounds unusually human. Its expressive qualities include laughter, pauses, and crying, features that may make an AI voice feel more like a conversational partner than a conventional text-to-speech system.

Block says Dessa was trained entirely on human speech and on a single NVIDIA GPU rack. The description highlights both the model’s training approach and the hardware used to develop it. The company has not announced public access to Dessa, so outside users cannot currently evaluate its performance directly. Its claims and demonstrations therefore provide the available picture of what the system can do.

Dessa is described as a full-duplex model. In practical terms, this means it can listen and speak at the same time, rather than requiring a strict turn-taking exchange in which one person finishes speaking before the system responds. The source compares this capability with OpenAI’s GPT-Live model, which OpenAI released earlier this year. Simultaneous listening and speaking could support interactions that feel more immediate, although the exact experience will depend on how the system handles interruptions, timing, and conversational context.

Many speech systems are paired with a large language model to retrieve information or produce responses based on a broader knowledge base. Block says Dessa’s knowledge comes from the conversational data it has. That distinction matters: a voice model’s ability to speak naturally is not identical to its ability to answer questions accurately. A system may convey tone and conversational flow while still depending on the information and capabilities available to it.

Block argues that existing voice AI can feel unnatural or irritating because text is placed in the process that drives speech generation. In many systems, a written response is generated first and then converted into spoken audio. That pipeline can preserve the words while losing some of the meaning conveyed through timing, tone, and other vocal cues. The company’s broader point is that speech is not merely text delivered aloud: how a person sounds can carry information that words alone do not express.

Some models, including ElevenLabs v4 and Fish Audio Drama 3, use voice tags to guide emotional delivery. Block’s claim is that emotion is built into Dessa itself. However, it is not clear from the available information whether the model’s emotional qualities are influenced by some form of tags or another steering mechanism. The distinction has not been fully explained, so the company’s statement should be understood as a description of its design approach rather than a complete account of how emotion is generated.

In a blog post, Block described its concerns about voice AI interactions: “Interacting with current voice AI can just plain suck. It turns your voice into text and throws away an incredible amount of information text alone can’t carry. The most powerful AIs in the world can’t hear how you feel. We think that’s ridiculous, leads to a worse experience, and is potentially dangerous.” This argument places emotional and vocal context at the center of the company’s case for developing a different kind of voice system.

Researchers at Block posted several examples of conversations with Dessa. In these demonstrations, the agent used different voices, cadences, and emotional tones. One example involved the bot saying that it was tired from work; at that point, its voice sounded tired as well. Another showed the bot speaking about a person making it comfortable, with an emotional delivery that sounded close to breaking down. A further example featured a Block team member playing a game with the voice assistant. These clips are intended to illustrate the range of conversational behavior the model can produce, although demonstrations alone do not establish how consistently the system performs across everyday situations.

Block says Dessa can run locally on a consumer-grade computer such as a MacBook M4 Max or better. Local execution may reduce the need to send every spoken interaction to remote computing systems. It could also make voice features easier to integrate into products and services, subject to the model’s hardware requirements and the practical details of deployment. The source does not provide further specifications about performance, latency, or the full range of compatible devices.

Despite the demonstrations, Dessa is not publicly accessible at the moment. Block has instead called for safety researchers to test the model. One likely focus is whether a voice agent could mislead or trick people on the other end of a conversation. Humanlike speech may make interactions easier and more expressive, but the same qualities could make it harder for listeners to recognize that they are speaking with an AI. Testing for deceptive or confusing behavior is therefore central to assessing the technology responsibly.

That concern is particularly relevant to Block’s business. The company operates in finance and sells solutions to small businesses. Square already uses voice AI to handle ordering, giving Block an existing connection to speech-based customer interactions. A proprietary model that runs on a device could potentially support more business and consumer experiences, while lowering costs compared with approaches that rely on external computing. These are possible strategic advantages, not confirmed product plans or quantified savings.

Overall, Dessa combines several ambitions: natural-sounding speech, expressive emotion, simultaneous listening and speaking, and local operation on consumer-grade hardware. Block presents these features as a response to the limits of systems that convert speech into text and then generate spoken output. But important questions remain about the model’s underlying emotional controls, the scope of its knowledge, its reliability in real use, and how it can be safeguarded against misuse. For now, Dessa remains a research model without public access, and Block is seeking safety researchers to help evaluate it.

Key Insights Table

Afterwards...

Dessa points toward a future in which voice interfaces may preserve more of the emotional and conversational context that can disappear when speech is reduced to text. Whether that approach improves everyday interactions will depend not only on how human the voice sounds, but also on the accuracy, transparency, and safeguards of the system.

Block’s next important step is evaluation. Safety researchers can help identify situations in which an expressive voice may confuse users, create misplaced trust, or be used to mislead someone. At the same time, further testing can clarify how well the model works on local hardware and whether its conversational behavior is dependable outside curated demonstrations.

If Block can address those challenges, locally run voice AI could become useful across a wider range of services, including small-business tools and consumer applications. For now, Dessa is a promising but not publicly available research effort, and its real-world value will become clearer only through broader testing and responsible deployment.