Block Unveils Dessa, a Human-Like Full-Duplex Voice AI Model

Image: TechCrunch · Source

Block introduces Dessa, a full-duplex voice AI that mimics natural human speech with emotional nuance and runs locally on consumer-grade hardware like MacBook M4 Max.

Block, the technology company co-founded by Jack Dorsey, has introduced a new voice model named Dessa through its research lab. Unlike many existing speech models that convert voice to text and lose much emotional nuance, Dessa operates fully on speech, enabling it to sound startlingly human with authentic expressions such as laughter, pauses, and even cries. The company describes Dessa as the most authentic full-duplex speech model ever built, capable of speaking and listening simultaneously, much like OpenAI's GPT-Live model.

Dessa’s design allows it to convey emotion and natural cadence during conversations, as the model is trained entirely on human speech using a single NVIDIA GPU. This contrasts with many current speech AI systems that rely on text intermediaries and voice tags to inject emotion, which can result in unnatural or monotonous outputs. Block asserts that emotion is inherently integrated into Dessa's speech generation, though it has not clarified whether specific emotion steering mechanisms like tags are used.

Examples shared by the company showcase the model expressing varying emotional states; in one instance, Dessa sounds tired while mentioning fatigue, and in another, it expresses vulnerability in response to human comfort. These interactions highlight the model’s capability to adapt cadence, tone, and emotional depth dynamically. Additionally, gameplay demonstrations reveal Dessa’s interactive capabilities in real-time, emphasizing its full-duplex functionality.

Significantly, Dessa can run locally on consumer-grade hardware such as the MacBook M4 Max, which promises broader accessibility without reliance on cloud computation. This local operation not only improves responsiveness but also enhances privacy and reduces operational costs. However, as of now, Dessa is not publicly available. Block is engaging with safety researchers to evaluate the model thoroughly, signaling caution especially concerning ethical implications like potential misuse to deceive users.

While Block currently operates largely in finance and small business solutions—using voice AI for tasks like ordering—owning a sophisticated voice model like Dessa could expand its offerings substantially. On-device voice AI could empower new business and consumer applications with richer, more human-like interactions at a lower cost. As Block continues to refine Dessa and address safety, its development marks a potentially significant leap forward in voice AI technology by integrating natural human expressiveness directly within the AI’s speech output.