Block unveils Dessa, a speech-only voice model designed for local use
The AMW Read
Block adds a speech-only, locally runnable model to the voice-model landscape, but restricted access and unverified performance limit its demonstrated impact.
Block unveils Dessa, a speech-only voice model designed for local use
Block has unveiled Dessa, a full-duplex speech-to-speech model that can listen and speak at the same time. Block says it trained the model entirely on human speech using a single NVIDIA GPU rack. Published demonstrations show voices with laughter, pauses, tired delivery, and emotional changes. The company says Dessa can run locally on a MacBook M4 Max or better. Public access is not available, and Block is seeking safety researchers to test it.
Dessa places Block in the voice-model supplier landscape alongside its existing financial-services business. The technical distinction is its speech-only approach: Block says text-based pipelines discard information carried by vocal expression. Dessa draws its knowledge from conversational training data, rather than the separate language model typically paired with speech systems. That makes conversational delivery and knowledge coverage separate questions for buyers. The demonstrations illustrate expressive speech, but the article provides no independent evaluation of accuracy, reliability, or operating costs.
For builders and investors, the concrete test is whether local, full-duplex speech can support dependable business interactions. Square already uses voice AI for ordering, giving Block a relevant application context. The article does not report a Dessa deployment in that workflow. Evaluation should test interruptions, factual answers, and whether emotional delivery misleads users about the system's identity. Claimed consumer-hardware support creates a deployment option; it does not yet establish lower costs or production readiness.