Block has introduced Dessa, a full-duplex speech-to-speech AI model trained entirely on human conversation audio without using any text or pretrained models. Revealed on October 23, 2026, as a research preview, Dessa is designed to listen and talk simultaneously, enabling natural interruptions, backchannels, and mid-sentence laughs. Unlike traditional voice agents that chain speech recognition, text models, and speech synthesis, Dessa processes audio directly to preserve emotion, sarcasm, and laughter.
The model was trained on a single NVIDIA GB200 NVL72 rack containing 72 Blackwell GPUs, making it significantly more data-efficient than text-first systems or competing models like Kyutai's Moshi. During inference, Dessa runs faster than real time locally on an Apple MacBook Pro M4 Max or better. Block has not yet published benchmark scores, parameter counts, weights, an API, or pricing details, though interested users can currently sign up for a waitlist on the project page.
This speech-only training approach matters because it bypasses the text bottleneck of conventional voice assistants, enabling more human-like, real-time conversational dynamics directly from audio signals.