Fintech company Block has introduced Dessa, a full-duplex speech-to-speech AI model trained from scratch using human conversation, without text data or pretrained models, that helps users have “human-like” conversation with the model.
According to Block, the model is designed to enable more natural, real-time conversations by recognizing vocal cues such as emotion, sarcasm, laughter and pauses. Unlike conventional voice AI systems that convert speech into text before processing it, Dessa directly processes and generates audio.
We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence.
https://t.co/r0qFLdvd6c
— jack (@jack) October 23, 2026
According to Block, Dessa is the first AI model it knows of that can support realistic live conversations after being trained exclusively on human speech.
“Dessa is the world’s first AI you can have a realistic live conversation with that was trained only from human speech. Dessa advances the frontier of deep learning models and creates a different category with entirely new and interesting behaviors for researchers to discover,” the company said.
The company said the model can listen and speak simultaneously , allowing it to interrupt naturally, respond to pauses and use conversational cues such as brief acknowledgments.
Block highlighted that training a model exclusively on speech presents significant challenges because spoken language contains information that text cannot capture, including variations in tone, accent, breathing and infection. The same sentence can convey different meanings depending on which words a speaker emphasizes.
The company also said Dessa learned language using hundreds of times fewer words than large language models typically require, although it acknowledged that the comparison does not account for the additional technical knowledge, including programming languages, that text-based models learn.
Block said it trained Dessa using a single NVIDIA GB200 NVL72 system, highlighting the computing setup used to develop the model. The project is intended to explore whether AI models can learn language and facts, acquire new languages and perform in-context learning efficiently through speech alone.
The company said the work could also improve how AI systems interpret human emotions during interactions. Block cited safety as one of the motivations behind the project, arguing that voice AI systems that fail to recognize distress or pain may respond inappropriately. Dessa is intended to help address this limitation by interpreting emotional cues directly from speech.
Dessa was developed by a small AI research lab within Block, founded by former members of Toronto based deep learning startup Dessa.
Also Read: Block CEO Dorsey Fires 4000 Staff Due to AI Even After Significant Profit
