Block’s new voice model sounds eerily human

Voice AI now sounds emotionally authentic and can run on your own hardware.

Block, Jack Dorsey's financial technology company, has unveiled Dessa, a new voice model developed by its research lab. Dessa is designed to sound remarkably human, incorporating natural elements like laughs, pauses, and cries, and is trained on human speech using a single NVIDIA GPU. It functions as a full-duplex model, enabling simultaneous talking and listening, akin to OpenAI's GPT-Live. Unlike many current models that rely on text-to-speech pipelines, Dessa's conversational data forms its knowledge base, aiming to deliver a more authentic and emotionally nuanced interaction. The company claims this approach overcomes the limitations of existing voice AI, which often discard crucial non-verbal information present in human speech. Dessa can reportedly run locally on consumer-grade hardware, such as a MacBook M4 Max, though public access is not yet available, with Block seeking safety researchers to test its capabilities.

Dessa's development signals a significant leap in natural language interaction, moving beyond mere text-to-speech to a more empathetic and human-like communication. By training on conversational data and enabling full-duplex communication, Block aims to create voice agents that are not just informative but also emotionally resonant. This focus on emotional nuance and real-time interaction could redefine user experiences across various applications, from customer service to personal assistants, making AI interactions feel more intuitive and less robotic.

The ability for Dessa to run on consumer GPUs marks a pivotal shift towards decentralized and on-device AI, potentially reducing reliance on cloud infrastructure and lowering operational costs for businesses. This decentralization could also enhance user privacy and enable more sophisticated voice-powered applications for small businesses, a key market for Block's Square services. As Block calls for safety researchers, the industry will be watching closely for potential misuse cases and the subsequent development of ethical guidelines for emotionally advanced voice AI.