Block’s Dessa voice model runs locally and sounds human

Block has released Dessa, a voice model from its research lab that runs locally and is designed to produce humanlike speech, including laughter, pauses and cries. The company says it was trained entirely on human speech using a single NVIDIA GPU.

Dessa is a full-duplex speech-to-speech model, meaning it can listen and speak at the same time. Block says its knowledge comes from conversational data used to train it. Researchers shared examples showing the model speaking with different voices, cadences and emotions, including a tired-sounding response and an emotional exchange.

Block says Dessa can run on consumer-grade computers such as a MacBook M4 Max or better. The model is not publicly available. The company has invited safety researchers to test it, including for cases where a voice agent might deceive someone.

Block says many existing voice models generate speech from text, which it argues can discard information conveyed by a speaker’s voice. The source article notes that it is unclear whether Dessa’s built-in emotion can also be steered with tags. Block already uses voice AI for ordering through Square, and an on-device model could support more business and consumer experiences, according to the article.