Block introduces local, full-duplex Dessa speech model

Block has introduced Dessa, a full-duplex speech-to-speech model from its research lab that the company says can run locally on a single NVIDIA GPU. The model listens and speaks at the same time, and Block says it was trained entirely on human speech rather than on a text-led speech-generation pipeline.

The company presented Dessa as a voice system intended to sound more natural through pauses, laughter, crying, emotional delivery and changes in cadence. Block says its conversational demonstrations include voices that sound tired when discussing work, emotional when describing comfort, and responsive during a guessing game.

A speech-first approach to voice interaction

Block’s central claim is that conventional voice AI loses important information when it converts spoken input into text before generating a reply. In its blog post, the company argued that text alone cannot retain how a speaker feels, and that this can lead to an inferior or potentially dangerous interaction.

Many voice-generation products use explicit voice tags to shape emotional delivery. The source cites ElevenLabs v4 and Fish Audio Drama 3 as examples. Block says emotion is built into Dessa itself, but it has not made clear whether that behaviour is also guided by tags or another steering method.

Like OpenAI’s GPT-Live model, Dessa is described as full duplex: it can process listening and speaking concurrently rather than treating a conversation as a sequence of separate turns. Block says the model’s knowledge comes from the conversational data on which it was trained. Speech models are commonly paired with a large language model for knowledge retrieval.

Local execution and safety testing

Block says Dessa can run on consumer hardware, naming a MacBook M4 Max or better as an example. That local capability could matter for applications where developers want speech processing to happen on-device rather than depend entirely on remote infrastructure.

There is no public access to Dessa at present. Instead, Block has called for safety researchers to test the system. The source notes that this could include cases in which a voice agent might deceive people on the other side of a conversation.

Potential relevance to Block’s business

Block operates in financial services and sells products to small businesses. Its Square business already uses voice AI for ordering. A proprietary model that runs on-device could potentially support additional business and consumer experiences at lower cost, as described in the source.

For businesses evaluating voice AI, Dessa’s announcement is a reminder to assess more than realism: local hardware requirements, the handling of conversational data, access controls and tests for deceptive behaviour should be considered before a highly expressive voice agent is put into customer-facing use.