Block unveils Dessa a voice AI model trained only on human speech

A shift toward audio-first AI

Jack Dorsey’s company, Block, has introduced a new speech model called Dessa. The company describes it as a full-duplex model, which means it can listen and talk at the same time. This allows for a more natural flow of conversation compared to traditional systems.

Most current voice AI systems work by turning spoken words into text first. This process often strips away important information like tone, emotion, and specific inflections. Block claims that these limitations can lead to a worse user experience and may even be potentially dangerous.

Dessa was built to solve these issues by focusing on the nuances of human speech. According to the company, the model is full of emotion, pauses, and a natural cadence. It aims to make talking with an AI feel like talking with a real person.

Training without text

One of the most distinct features of Dessa is its training methodology. While most large language models rely on massive amounts of text, Dessa was trained from scratch using only real human conversations. The developers did not use any text data or pretrained models during the process.

The company says the model learned to hold a conversation simply by listening to audio. This required the AI to study complex elements like accents, breathing patterns, humor, and sarcasm. It had to learn not just what words mean, but also the specific timing of when to stop speaking.

The technical challenge of this approach is significant because a single spoken word contains much more data than a written one. Block claims they cracked the code by training the model on a single NVIDIA GB200 NVL72. This achievement allows the model to be compact enough to run on a laptop.

Real-time human interaction

The model is designed to handle the unpredictability of live speech. In public demos, Dessa is shown filling in when a user blanks out mid-sentence. It can react with a gasp or a laugh that feels genuinely empathetic rather than a programmed response.

Because there is no transcription layer in between, the inference runs faster than real time. This means there is no latency, which is the delay between a user speaking and the AI responding. The model autonomously picks what to say and how to say it with its own distinct personality.

The company even demonstrated Dessa interacting with ChatGPT Live. These interactions highlight the model's ability to maintain a realistic live conversation. It can laugh casually mid-sentence if something funny happens, rather than waiting for a user to finish before reacting.

Research and future questions

Dessa was created by a small AI research lab at Block founded by former members of a Toronto deep learning startup. The researchers have spent years building generative speech models and deep learning systems. They view this new model as an advancement that offers new avenues for discovery.

The launch raises several questions for the research community regarding how AI learns. Experts want to know how efficiently a model can learn language and facts when it is not provided with text. There are also questions about how efficiently it can learn new languages or perform in-context learning.

Currently, Dessa is only available in a research preview behind a waitlist. The company has not yet released a parameter count, API, price, weights, or published evaluations. For now, the public can only see the capabilities of the model through the demos provided by Block.

The philosophy of natural interaction

Jack Dorsey shared the news of the model's development on his X account, highlighting his belief that voice is the most natural way for people to interact with intelligence. He described Dessa as the most authentic full-duplex speech model and expressed his pride in the team for pushing the frontier of what is possible in AI flexibility.

The company argues that current voice AI is limited because it converts speech into text, a process that overlooks a massive amount of information. Because text is identical regardless of the author, it cannot capture the unique way a person speaks. The official website claims that most powerful models today cannot hear how a user feels, which can lead to a worse experience.