Block Unveils Dessa, a Speech-Native Full-Duplex AI Voice Model

Financial technology company Block has introduced Dessa, a full-duplex speech model designed to deliver real-time, authentic spoken conversation. Developed by Block's internal artificial intelligence research lab under the leadership of Ragavan Thurairatnam, the project marks a significant shift away from traditional voice assistant pipelines. The announcement, published on the company's dedicated portal, frames the system as an effort to recreate human conversational dynamics from the ground up.

Unlike dominant voice systems that rely on multi-step pipelines, Dessa was trained from scratch entirely on human conversational audio. Conventional voice assistants typically transcribe incoming audio into text, feed that transcript through a standard large language model, and then synthesize the generated text response back into speech. This layered approach often introduces perceptible latency and strips away expressive conversational cues before the model can respond.

By contrast, Dessa bypasses written text entirely, omitting traditional datasets such as digitized books, articles, and scraped web text. The development team designed the training process to mirror how humans naturally acquire language during early childhood by listening to vocal patterns, rhythm, and cadence. As a result, the model processes spoken language directly as raw sound rather than abstract textual tokens.

A central capability of the new model is its full-duplex architecture, which enables simultaneous listening and speaking. Most automated voice interfaces operate in half-duplex mode, functioning much like a walkie-talkie where one participant must fall silent before the other can transmit. Dessa is built to behave like an open phone line, continuously tracking the audio feed while generating its own voice output in parallel.

This continuous processing allows the system to support natural human conversational quirks that often break conventional conversational bots. Dessa can handle spontaneous interruptions, mid-sentence topic shifts, and backchanneling cues such as vocal affirmations and laughter. It also registers non-verbal vocal dynamics, including pauses, sighs, emotional tone adjustments, and conversational pacing, without requiring discrete command prompts.

The project draws its name from Dessa, the Toronto-based artificial intelligence firm acquired by Jack Dorsey's company in 2020. Over the years following that acquisition, Block integrated the team's machine learning expertise into internal operations while maintaining exploratory research into applied intelligence. The release of this speech model represents one of the lab's most visible public demonstrations of fundamental model engineering to date.

In addition to its speech-first architecture, Block emphasized that Dessa was engineered with operational efficiency in mind. Rather than relying entirely on remote cloud clusters and high-bandwidth streaming connections, the model is compact enough to run locally on high-end consumer hardware. Company demonstrations indicated that the software can execute locally on machines such as an Apple MacBook Pro equipped with an M4 Max processor.

The transition toward native audio-to-audio modeling reflects a growing consensus among artificial intelligence researchers that text bottlenecks limit interactive realism. By remaining in the audio domain, speech-native systems reduce the processing lag that makes automated interactions feel disjointed. However, training without large text corpuses also raises practical questions regarding the depth of factual knowledge the model can retain without supplementary text retrieval.

Block has made demonstrations available as industry observers evaluate how speech-first models will integrate into real-world software products. As conversational interfaces continue to evolve beyond rigid prompt-and-response paradigms, architectures like Dessa suggest that the next generation of voice tools will prioritize natural cadence, local deployment, and real-time responsiveness over traditional text-mediated pipelines.