In brief: Block has introduced Dessa, a local, human-sounding full-duplex speech-to-speech model trained on conversational data, currently inviting safety testing.
Block has introduced a new speech-to-speech voice model called Dessa, developed within its research lab. The technology stands out for producing sound that is remarkably human-like while operating locally rather than requiring constant cloud-based processing infrastructure.
Capabilities and Hardware Requirements
Dessa functions as a full-duplex model, meaning it has the capability to talk and listen simultaneously during an interaction. Unlike many standard conversational systems that rely entirely on massive remote data centers, this system is designed to run locally. It can execute operations on a single NVIDIA GPU rack or on higher-end consumer hardware, specifically a MacBook M4 Max or better.
The underlying training data for the model consists entirely of human speech. Rather than depending on intermediate text steps to generate responses, Dessa derives its knowledge directly from conversational data. However, certain technical specifics regarding its emotional output remain uncertain, as it is currently unclear if the emotion present in the model's speech is actively steered by voice tags.
Integration and Current Access Restrictions
While the broader technology is new, practical implementation within Block's ecosystem is already underway. Square, a property under the Block umbrella, currently utilizes voice artificial intelligence to manage customer ordering processes.
Despite these internal applications, public access to Dessa is not available at this time. Instead of an open release, Block is actively calling for safety researchers to evaluate the model and test its capabilities before any wider rollout is considered.
Conclusion
Block's introduction of Dessa marks a notable step in local, conversational voice technology. By combining simultaneous listening and speaking capabilities with consumer-hardware compatibility, the project highlights ongoing developments in speech-to-speech models, though rigorous safety testing remains a primary current focus.
Source for this report: techcrunch.com.