Block has unveiled Dessa, a speech-to-speech AI model built from scratch entirely on human audio rather than text data. Trained on a single NVIDIA GB200 NVL72 and capable of running locally on a MacBook Pro M4 Max, Dessa processes audio directly to preserve emotional cues, sarcasm, and real-time pacing. Block is restricting early access to AI safety researchers to test for unexpected risks before considering a public release.
Dessa speech-to-speech AI model
- ▪Block, the company behind Square and Cash App, announced an AI speech model named Dessa
- ▪Block is withholding a public release for Dessa while offering early access to AI safety researchers for unexpected risk testing
- ▪Block developed Dessa as a speech-to-speech model that processes audio directly without converting speech into text
Training on human audio
- ▪Block trained the Dessa AI model using a single NVIDIA GB200 NVL72 unit
- ▪Block trained Dessa from scratch on human speech without using written text data
Technical specifications
- ▪Block designed Dessa to run locally on an Apple MacBook Pro M4 Max or higher system
- ▪Block's Dessa model supports full-duplex communication, enabling it to process speech and respond simultaneously
Emotional intelligence capabilities
- ▪Prior emotional voice AI initiatives include Hume AI's EVI launched in March 2024, Kyutai's Moshi in September 2024, and OpenAI's ChatGPT Advanced Voice Mode overhauled in June 2025
- ▪Dessa processes direct audio signals to capture emotional tones, sarcasm, laughter, and pauses in real time
Story comments
Loading comments…