Block‘s AI research lab has introduced Dessa, a new full-duplex speech-to-speech model. According to reporting by MarkTechPost, the system was trained exclusively on human conversation audio rather than traditional text transcripts. The model is capable of running locally on hardware such as a MacBook Pro M4 Max.
Understanding Full-Duplex Speech Models
Full-duplex audio systems allow for simultaneous two-way communication, meaning a user can speak and listen at the same time without the turn-taking delays typical of older voice assistants. Training models directly on raw speech audio bypasses the intermediate step of converting speech to text and back, aiming for more natural and fluid machine interactions.
Unknown Specifications and Missing Details
Despite the release, Block has not published the model size, performance benchmarks, or model weights. The announcement did not specify when or if these details will be made publicly available. Observers will need to wait for further disclosures from the company to properly evaluate Dessa’s capabilities against other speech models.
Further technical documentation from Block will be necessary to understand Dessa’s architecture and potential applications.
Lead image: Yusuf Çelik / Pexels
