Block Unveils Dessa, a Full-Duplex Speech-to-Speech AI Model

Block has unveiled Dessa, a native full-duplex speech model designed to conduct fluid, bidirectional voice conversations without relying on text transcription. Developed by an in-house artificial intelligence research unit at the Jack Dorsey-led technology and financial services firm, the model can listen and speak simultaneously in real time. The announcement marks a significant departure from standard conversational systems that depend on sequential turn-taking.

Most contemporary voice agents rely on a multi-stage pipeline that transcribes speech to text, passes the tokens to an underlying large language model, and converts the resulting text back into synthetic audio. In contrast, Dessa functions entirely within the audio domain from input to output. Block disclosed that the architecture was trained from scratch exclusively on raw human speech and environmental audio, without utilizing written text, digital books, or scraped web documents.

By removing the intermediary text layer, the system preserves acoustic and emotional nuances that conventional speech recognition pipelines routinely discard. The model natively captures inflections, hesitations, conversational sighs, subtle laughter, and tonal shifts such as sarcasm or distress. As a result, Dessa generates vocal responses that mirror natural human cadence and expressiveness rather than synthetic recitation.

The core breakthrough highlighted by Block centers on the model's authentic full-duplex operation, which mimics the natural flow of a telephone call rather than a half-duplex walkie-talkie. Traditional voice interfaces must wait for a speaker to stop talking before processing a reply, frequently misinterpreting natural pauses as conversational handoffs. Dessa continuously processes incoming audio while generating speech, allowing users to interrupt freely, interject thoughts, or receive spontaneous conversational affirmations such as brief acknowledgments.

In addition to its novel training approach, Block emphasized that Dessa achieves low-latency inference speeds that outpace real-time requirements. The engineering team designed the model to operate locally on modern consumer workstations, demonstrating live execution on Apple silicon hardware including MacBook Pro systems powered by M4 Max processors. Running locally eliminates cloud transmission latency while providing enhanced user privacy and offline operational stability.

The project was developed under the leadership of Ragavan Thurairatnam, a veteran machine learning researcher who directs Block's specialized artificial intelligence lab. Thurairatnam originally co-founded the Toronto-based machine learning company Dessa, which Square acquired in 2020 prior to its corporate rebranding as Block. The current research initiative operates as a compact, focused team reporting directly to Dorsey on frontier artificial intelligence systems.

The initiative reflects a broader strategic expansion for Block beyond its core financial ecosystems, which include merchant terminal services under Square and consumer banking utilities within Cash App. While financial technology remains the company's economic foundation, leadership has increasingly prioritized in-house deep learning research and decentralized computing tools. Developing proprietary audio architecture provides Block with specialized technological autonomy as artificial intelligence infrastructure becomes increasingly proprietary.

The debut of Dessa arrives as competition intensifies across the artificial intelligence sector to master real-time multimodal interaction and human-grade voice response. While industry leaders have released advanced speech modes, many still grapple with conversational interruptions, robotic latency, and emotional flattening caused by text bottlenecks. Native audio-to-audio foundation models represent an active research frontier that seeks to bypass symbolic text processing altogether.

Block has published interactive audio demonstrations and architectural details on its dedicated web portal, offering an early glimpse into the technology through restricted developer access. Moving forward, the research team is expected to refine conversational stability, evaluate broader hardware compatibility, and explore practical integration pathways across ambient audio devices and decentralized consumer applications.