Block unveils Dessa voice AI that mimics human emotion and cadence

A new standard for speech

Block, the company behind Square and Cash App, has introduced a new voice AI model called Dessa. Unlike many existing systems that rely on text-to-speech pipelines, this model was trained exclusively on human speech. This training method allows the AI to capture the nuances of human communication more effectively than models built on written data.

The technology is designed to handle the messy realities of human conversation. It can manage interruptions, react to laughter, and adjust its tone based on the emotional context of a dialogue. Because it processes audio directly, it avoids the limitations of systems that must first translate spoken words into text before generating a response.

Jack Dorsey, the co-founder of Block, highlighted the realism of the model on social media. He noted that interacting with the AI feels like talking to a person because it includes natural pauses and a distinct cadence. The company demonstrated the model reacting to surprises with what appeared to be genuine emotion during live recordings.

Technical architecture and hardware

Block utilized a single Nvidia GB200 NVL72 computing system to train the model from scratch. Despite the intensive training process, the resulting model is remarkably compact. This efficiency means it can run locally on a MacBook Pro equipped with an M4 Max chip or better, rather than requiring massive cloud infrastructure for every interaction.

The development of Dessa involved a specialized AI research lab within Block. This team includes former staff from a Toronto-based startup called Dessa, which Block acquired in 2020. The lab reports directly to Dorsey and focuses on pushing the boundaries of how machines interpret and generate human-like audio.

While the company has shared several demonstrations, it has not yet released specific technical benchmarks or the model's size. The current status of the project is a research preview. This means the technology is not yet available to the general public and is being shared selectively with experts.

Mixed reactions from the community

The debut of Dessa has sparked a polarized reaction across social media. Some users have expressed awe at the level of realism, describing the interactions as feeling like science fiction. These users point to the AI's ability to mimic human emotion as a significant leap forward in interactive technology.

Conversely, the same humanlike qualities have raised alarms among security watchers and filmmakers. Critics argue that if an AI can sound indistinguishable from a human, it becomes a potent tool for deception. There are concerns that such technology could be used to trick people into lowering their guard during sensitive conversations.

Some observers have specifically warned about the potential for a rise in sophisticated fraud. They suggest that high-quality voice cloning or generation could lead to a surge in scam calls. These concerns highlight the tension between the desire for more natural AI and the need to protect users from malicious exploitation.

Potential uses and safety risks

Despite the fears of fraud, there are significant opportunities for the technology in specialized fields. For example, Binance founder Changpeng Zhao suggested that natural voice interactions could be highly beneficial for children's education. He noted that teaching kids through conversation is a challenge when using current, less natural AI voices.

Block suggests that speech-based AI could eventually become much better at recognizing emotional cues. This includes identifying signs of distress or pain, which are often lost in standard text transcripts. By processing the audio directly, the model may be able to provide more empathetic or accurate responses in healthcare or support settings.

To address these possibilities and risks, Block is offering early access to a small group of AI safety researchers. These experts are tasked with assessing the dangers of the model before it reaches a wider audience. The goal is to identify any overlooked risks while exploring how the technology can be used responsibly.