Block unveils Dessa, an on-device full-duplex speech model engineered for natural conversational nuance
Jack Dorsey’s company debuts an end-to-end audio model running locally without cloud round-trips, delivering human-like laughter, pauses, and cadence.
Technology news
Jack Dorsey’s digital infrastructure and payments firm Block has released a breakthrough voice model developed by its internal research division. Dubbed Dessa, the model operates natively on consumer hardware and replicates human speech patterns with uncanny realism, including unprompted pauses, emotive laughter, breath catch, and natural conversational cadence.
Unlike standard voice assistant setups that daisy-chain a speech-to-text pipeline into an external cloud-hosted large language model and subsequently pass generated text to a speech synthesizer, Dessa functions as a native, full-duplex audio architecture. This allows the system to simultaneously listen and speak, permitting real-time human interruptions and conversational overlapping without perceptual lag.
Block confirmed that Dessa was trained exclusively on human dialogue data utilizing an efficient single Nvidia GPU rack, bypassing the multi-megawatt computing clusters typically required for frontier conversational models. Its underlying contextual knowledge is embedded directly within its conversational weights rather than queried from external document retrieval layers.
By keeping speech processing fully on-device, Block intends to deploy Dessa across hardware touchpoints and point-of-sale interfaces without transmitting raw customer audio to cloud servers. The announcement reflects a growing industry pivot toward compact, low-latency audio models capable of fluid dialogue without sacrificing user privacy.