A small research team at Block has developed a speech-to-speech AI model capable of simulating lifelike human interaction. The new model, called Dessa, learned to hold real-time conversations by listening to human speech alone, without the use of text or pretrained models. The company says the system was trained on a single GB200 NVL72 GPU and is small enough to operate on a MacBook.
Designed as a full-duplex model, Dessa incorporates pauses, emotion and natural cadence to remove the artificiality often found in AI voice assistants. Block is currently granting access via a waitlist and has invited AI safety researchers to test the model’s capabilities, though the tool is not yet available for general public use.