Jack Dorsey’s Block has unveiled Dessa, an experimental speech-to-speech AI model designed to communicate more naturally by learning directly from human conversations rather than written text.
Block, the financial technology company behind Square and Cash App, has introduced Dessa, a new artificial intelligence model designed to make voice conversations feel more natural and human-like.
Unlike many voice assistants that process speech by converting it into text before generating a response, Dessa is a full-duplex speech-to-speech model. It can listen and speak simultaneously, allowing it to interrupt, acknowledge comments and respond without waiting for a speaker to finish.
According to Block, Dessa was trained from scratch using recordings of human conversations. The company says the model does not rely on written text or pretrained language models, with its conversational abilities developed directly from audio.
AI Designed to Understand Human Conversation
Block says training Dessa on audio alone could help it recognise conversational cues that conventional speech-to-text systems may overlook, including sarcasm, hesitation, laughter and changes in emotion.
Pakistan Elect to Field Against England in Crucial ODI Tri-Series Clash
Demonstrations published on the company’s website show the model laughing during a sentence, adjusting its speaking volume when requested and helping complete a sentence when a person forgets a movie title.
These features are intended to make conversations feel less like interactions with a conventional digital assistant and more like spontaneous exchanges between people.
One demonstration features Dessa interacting with OpenAI’s GPT-Live-1 voice model. Block says the recordings show live interactions without anyone guiding the responses. However, these are company-produced demonstrations and should be considered alongside independent testing as it becomes available.
Trained on Nvidia’s Advanced AI Hardware
The development of Dessa also highlights the computing resources required to build advanced speech models.
Block says it trained the model on a single Nvidia GB200 NVL72 system equipped with 72 Blackwell GPUs.
Despite that training setup, the company reports that the completed model can run on a MacBook Pro equipped with an Apple M4 Max chip or a more powerful configuration.
Block also says Dessa operates faster than real time, potentially allowing it to maintain fluid conversations without noticeable delays.
The company has not yet disclosed the model’s parameter count, released its model weights or published standard benchmark results, leaving its performance difficult to assess independently.
Limited Testing Before Public Release
Dessa is not yet publicly available. Block is inviting a small group of AI safety researchers to test the model, while other interested users can join a waiting list.
The early testing phase could help researchers examine the model’s conversational capabilities, limitations and potential safety concerns before broader access is considered.
For now, the project represents an alternative approach to voice AI: rather than building conversations around text-based language processing, Dessa attempts to learn the patterns of human communication directly from audio.
Whether that approach delivers more reliable, emotionally aware and natural interactions will become clearer as independent researchers gain access to the system.
