Jack Dorsey’s Block unveils Dessa, an AI voice model trained entirely on human conversations

Jack Dorsey’s Block has introduced Dessa, a new AI voice model trained entirely on human conversations rather than written text, designed to communicate in real time while recognizing emotional cues, interruptions and natural speech patterns.

The company said Dessa was trained from scratch using only recordings of human speech, without text data or pretrained models.

Unlike conventional voice assistants that convert speech into text before generating a response, Dessa processes and generates audio directly through a speech to speech architecture.

The model supports simultaneous listening and speaking, allowing it to interrupt, respond to pauses and adjust its tone during conversations.

Block said Dessa can run locally on a MacBook Pro equipped with Apple’s M4 Max chip or better, without requiring cloud infrastructure for inference.

The company has published several demonstrations of the model and plans to provide early access to a small group of AI safety researchers.

Dessa learns to communicate without text training

According to Block, Dessa was developed to explore whether AI models could acquire language and conversational abilities through speech alone, similar to how humans learn to communicate before learning to read.

“Babies learn to talk without ever seeing a word written down. We wanted to see if a model could too. It can. It’s just very hard,” the company said.

The model uses what Block describes as a full duplex architecture, meaning it can listen and speak simultaneously rather than waiting for a person to finish before responding.

This allows Dessa to recognize laughter, sarcasm, changes in tone and emotional expressions that can be lost when speech is converted into text.

Block said the model learned language using hundreds of times fewer words than conventional text based language models, although it did not provide a detailed comparison.

The company also said Dessa was trained using a single NVIDIA GB200 NVL72 system, suggesting that speech focused AI development may require a different approach to computing infrastructure.

In demonstrations shared by Dessa lead Ragavan Thurairatnam, the model engages in conversations about movies, responds to personal stories and interacts with OpenAI’s GPT Live 1 voice model.

Other examples show Dessa participating in guessing games and adjusting its responses during conversations.

However, some demonstrations also reveal limitations, including instances where the model misidentifies a speaker or unexpectedly changes topics.

The examples were selected and published by Block, and the model’s performance has not been independently evaluated.

Block targets emotional awareness and AI safety

Beyond conversational capabilities, Block said Dessa was developed to investigate how AI systems could better understand human emotions and respond appropriately in sensitive situations.

The company argues that traditional voice assistants lose important information when converting speech into text, including signs of frustration, distress and emotional changes.

“AI keeps gaining power and independence, and it’s still nearly deaf to the emotion in our voices,” the Dessa team said.

Block pointed to situations where a person experiencing serious pain might interact with an AI system that fails to recognize their distress.

The company said its approach could help address those limitations by allowing AI models to interpret emotional signals directly from speech.

The project is being developed by a small research team at Block led by Ragavan Thurairatnam, with members who previously worked at the Toronto AI startup Dessa.

The researchers report directly to Dorsey and have spent several years developing generative speech technologies.

Block plans to grant early access to selected AI safety researchers to identify potential risks and study how models trained exclusively on speech behave.

The company has not disclosed a broader release date, model weights or detailed performance benchmarks.