Block Unveils Dessa AI Voice Model Trained on Human Speech
Block launched Dessa, a full-duplex AI voice model that processes audio directly to mimic natural human emotion and conversation without using text intermediaries.
Financial technology company Block has unveiled Dessa, a full-duplex speech-to-speech AI model designed to mimic natural human conversation. Unlike traditional voice AI that converts speech to text, Dessa processes audio directly, allowing it to listen and speak simultaneously. This capability enables the model to interrupt users and recognize emotional cues, including sarcasm, laughter, and hesitation.
Block developed the model from scratch using only recordings of human speech, avoiding pretrained language models. The training was conducted on a single Nvidia GB200 NVL72 system, and the company states the model is efficient enough to run locally on a MacBook Pro equipped with an M4 Max chip. Demos released by the company show Dessa interacting with humans and OpenAI's GPT-Live-1.
Currently, the model is a research preview and is not available to the general public. Block is granting early access to a small group of AI safety researchers to identify potential risks while other users remain on a waitlist. The launch has sparked mixed reactions, with some praising the authenticity of the interaction and others warning that such lifelike AI could increase social isolation or facilitate sophisticated scam calls.
