Jack Dorsey's Block Unveils Dessa, A Voice Model That Talks Like A Person
0:00 / 0:42
News
Jack Dorsey's Block Unveils Dessa, A Voice Model That Talks Like A Person
calendar_today Date:
schedule Duration: 0:42
database
Summary Report
Block's Dessa is a full-duplex speech-to-speech model trained from scratch on human conversation.
- 01. Listens and talks at once; keeps emotion, sarcasm, laughter and pauses
- 02. Trained on one GB200 system; runs locally on a MacBook Pro; waitlist only
Block has unveiled Dessa, which it calls the most authentic full-duplex speech model ever built. It listens and talks at the same time, dropping in a 'yeah' or 'got it' while you speak, and handles interruptions naturally. It works speech to speech with no text in between, keeping emotion, sarcasm, laughter and pauses. Block trained it from scratch on human conversation on a single NVIDIA GB200 system, and it runs locally on a high-end MacBook Pro. No benchmarks yet; early access is via waitlist and for AI safety researchers.
