Block Launches Dessa Full-Duplex Voice Model for Local Real-Time Interaction

Block has launched Dessa, a full-duplex AI voice model trained entirely from scratch on human speech data, eliminating the traditional "speech-to-text" intermediate step. Dessa directly processes acoustic signals to capture and generate real-time emotions, sarcasm, pauses, and tonal nuances, achieving native emotional expression that addresses the rigidity of conventional TTS technologies.

Cross-verified reports indicate that Dessa employs an end-to-end architecture supporting simultaneous listening and speaking, similar to OpenAI's GPT-Live but with a stronger focus on preserving paralinguistic information. Its knowledge derives directly from dialogue data without requiring external LLM retrieval. Notably, the model is highly efficient, capable of running locally on consumer devices like the MacBook Pro M4 Max or a single NVIDIA GPU, significantly lowering hardware barriers and privacy risks.

Currently restricted to security researchers for deception risk assessment, Dessa marks a paradigm shift from "transcription-based interaction" to "human-like real-time communication." This breakthrough offers a new path for low-cost edge deployment in commercial scenarios, potentially reshaping user experiences in customer service and companion applications.