Block releases Dessa voice AI that mimics human emotions and pauses

A shift toward natural speech

Jack Dorsey's Block has unveiled a new voice AI model called Dessa. The company designed this model to move away from the robotic feel of current machine-generated speech. It aims to create interactions that feel more like conversations between two people.

Dessa is capable of expressing a wide range of human emotions through its voice alone. The model can sound tired from work, appear close to tears, or laugh during a sentence. It also incorporates natural cadences such as pauses and changes in vocal emphasis.

According to the International Business Times Singapore, the goal is to make AI interactions feel more authentic. Block believes that voice is the most natural way for humans to interact with intelligence. They want the experience to include nuance and emotion rather than just functional responses.

How the audio pipeline works

Most current voice AI systems use a pipeline that converts your spoken words into text before generating a response. Block argues that this process strips away vital information like sarcasm, breathing, and emotional tone. They claim that text alone cannot carry the full weight of how a person feels.

Dessa takes a different approach by being a full-duplex, speech-to-speech model. This means it processes audio directly without an intermediate text step. By listening to human conversations during training, the model learned to conduct real-time dialogue naturally.

Because it processes audio directly, the system can recognize specific conversational cues. It can use backchannels like "yeah!" or "got it!" to show it is listening. It is also designed to know when it is appropriate to interrupt the user or respond to a shift in their tone.

Hardware and training details

The development of Dessa involved a specific training process using human speech data. Block trained the model from scratch rather than using pretrained models or text-based data. The entire training process was completed on a single NVIDIA GB200 NVL72 chip.

One of the notable technical features of Dessa is its ability to run locally. This means the AI does not have to rely on a constant connection to a massive cloud server to function. Specifically, the model can run on a MacBook Pro equipped with an M4 Max chip or better.

Running the model locally has significant implications for how companies deploy voice technology. It allows for the creation of consumer and business experiences without the high costs of cloud inference. This could make it easier for companies to integrate sophisticated voice features into various products.

Safety and early access

Dessa is not yet available to the general public for use. Block is currently providing early access to a small group of AI safety researchers. The company wants these experts to investigate the risks associated with a model that produces such convincing human-like speech.

Block stated that a model learning to talk from sound alone is new territory. They want outside eyes to find risks that the internal team may not have anticipated. This cautious rollout highlights the potential for the technology to be used in ways that require careful oversight.

The company is looking for feedback on how the model handles emotional delivery. While it is unclear if the system uses internal tags to steer its emotions, the company claims the ability is built into the system. The focus remains on understanding how these realistic interactions might affect human users.

Commercial and future impact

The release of Dessa could have significant commercial implications for Block's existing services. The company already operates financial and small-business services, and its subsidiary Square already uses voice AI for ordering. This new model could enhance those existing tools with more human-like interaction.

Beyond just improving customer service, the technology represents a shift in AI architecture. By making voice the foundation of a conversation rather than text, Block is changing how humans might interact with computers. It moves the needle toward a more immersive and emotionally aware digital experience.

As the technology matures, the focus will likely shift toward wider deployment. For now, the priority is ensuring the model is safe and reliable. The transition from cloud-dependent systems to local, high-fidelity voice models marks a new chapter in AI development.