Block’s Dessa: A Local Voice AI Model With Built-In Emotion
Block’s new Dessa model runs locally and captures human emotion without cloud dependency.
On-Device Architecture and Latency
Native Emotional Expression
Privacy and Safety Considerations
Block has unveiled Dessa, a voice AI model designed to run entirely on local hardware. Unlike cloud-dependent systems, Dessa processes speech on-device, enabling features like full-duplex communication and nuanced emotional expression such as laughter or pauses. This shift aims to solve the latency and privacy issues inherent in transmitting audio to remote servers.
The architecture diverges from standard speech-to-text pipelines. Instead of relying on a paired large language model for knowledge, Dessa draws from its conversational training data directly. Block claims this preserves emotional cues often discarded in text conversion. While competitors use tags to steer emotion, Dessa integrates these expressions natively, though the exact mechanics remain under review by safety researchers.
For developers, the ability to run on consumer hardware like a MacBook M4 Max suggests significant cost reductions. This could empower applications like Square’s ordering systems to scale without expensive GPU infrastructure. However, public access is limited as Block invites safety auditors to evaluate the model before wider release.
A promising step toward private, responsive voice interfaces, but wait for public access.