Block launches Dessa AI voice model trained only on human speech

A new approach to audio AI

Block, the financial technology company led by Jack Dorsey, recently introduced a new artificial intelligence voice model called Dessa. The company designed this model to move beyond the limitations of current voice systems that rely on converting speech into text first. By skipping that middle step, the AI aims to preserve the subtle nuances of human communication.

Dessa is what the company describes as a full-duplex speech-to-speech model. This means it is capable of listening to a person and speaking back at the same time, rather than waiting for a turn to finish. This capability allows for a more fluid and natural flow of conversation that mimics human interaction more closely.

The primary goal of this technology is to make interactions with intelligence feel more authentic. Block claims that talking with Dessa feels like talking with a person because it incorporates natural cadence, pauses, and emotional depth. The company believes that voice is the most natural way for humans to interact with AI technology.

Training on sound alone

One of the most distinct features of Dessa is how it was built. The company stated that they trained the model from scratch using nothing but human conversation sounds. They did not use any text data or existing pretrained models to teach the AI how to communicate. This is a significant departure from how many large language models are typically developed.

The company noted that training on speech is inherently complex because spoken words contain much more data than written text. Every utterance varies in tone and delivery, which presents a unique challenge for machine learning. However, Block claims that Dessa learned language from hundreds of times fewer words than typical large language models require.

Despite the complexity of processing raw audio, the company managed to train the model efficiently. They reported that they successfully figured out how to train Dessa on only one NVIDIA GB200 NVL72. This suggests a focus on optimizing the model's architecture to handle high-density audio data without requiring massive amounts of hardware.

Capturing human emotion

Because Dessa processes audio directly, it is able to recognize and express a wide range of human emotions. The model can identify and mirror feelings such as frustration, sadness, and even sarcasm. It can also react to humor by laughing in the middle of a sentence, which is a behavior it was not explicitly programmed to do.

The AI is also capable of adjusting its own physical delivery based on the context of the conversation. For example, it can change its speaking volume if a user asks it to do so. It can also respond spontaneously to various prompts, making the interaction feel less like a series of commands and more like a dynamic dialogue.

The company highlights that these behaviors emerge naturally from the model's training on human sounds. By listening to how people actually talk, the AI learned to navigate the nuances of tone and emotion. This allows it to maintain a consistent conversational style that feels more human-centric than traditional robotic voices.

Safety and future testing

Block acknowledges that creating a model that learns to talk from sound alone represents new territory in the AI field. Because the technology is so different from standard text-based models, the company wants to ensure it is handled responsibly. They are currently looking for outside perspectives to evaluate the system's behavior and potential risks.

To facilitate this, the company is providing a small group of AI safety researchers with early access to the model. These experts will work to identify any risks that the developers may not have anticipated during the initial build. This step is intended to provide a layer of scrutiny before the model sees wider use.

As the technology moves forward, the focus will likely remain on refining how these emotional cues are used in real-world applications. The goal is to create a model that feels authentic while remaining safe for public interaction. By involving researchers early, Block aims to set a standard for how full-duplex speech models are deployed.