Block has introduced a sophisticated new voice model named Dessa, developed within the company’s research division. Unlike many existing AI voice systems that rely on text-to-speech pipelines or external voice tags to simulate emotion, Dessa is trained entirely on human speech. This approach allows the model to naturally incorporate nuances such as laughter, pauses, and emotional inflection, creating a conversational experience that closely mimics human interaction.

One of the standout features of Dessa is its full-duplex capability, which enables the AI to listen and speak simultaneously. This functionality mirrors the real-time responsiveness seen in advanced models like OpenAI’s GPT-Live. By bypassing the traditional method of converting speech to text and back again, Block claims its model preserves the emotional context and subtext that are often lost in standard AI interactions, which the company argues can lead to disjointed or frustrating user experiences.

Technically, Dessa is designed for efficiency, capable of running locally on consumer-grade hardware, such as the MacBook M4 Max. This on-device capability offers significant potential for privacy and cost-effective deployment, particularly for businesses looking to integrate voice AI into customer service or point-of-sale systems. While the model is not yet available for public use, Block has initiated a safety research phase, inviting experts to evaluate the technology for potential risks, such as its ability to convincingly mimic human emotion in deceptive scenarios.

For Block, which already utilizes voice AI for its Square payment solutions, the development of an in-house, high-fidelity model marks a strategic shift. By controlling the underlying technology, the company aims to enhance its business and consumer offerings, moving toward a future where voice serves as the primary, most natural interface for human-AI interaction.

Key Takeaways

• Dessa is a full-duplex, speech-to-speech AI model trained entirely on human audio, allowing for natural emotional expression without the need for text-based tags.
• The model is optimized to run locally on consumer hardware, such as the MacBook M4 Max, reducing reliance on cloud-based processing.
• Block is currently restricting access to the model to prioritize safety research, specifically focusing on the potential for misuse in deceptive voice-based interactions.

Editor’s Analysis & Impact

The introduction of Dessa signals a significant evolution in the voice AI landscape, moving away from robotic, text-dependent synthesis toward emotive, context-aware communication. By enabling local execution on consumer hardware, Block is effectively democratizing high-end AI performance, which could disrupt the current cloud-dependent business models of major AI providers. From a market perspective, this technology has immense implications for the fintech and retail sectors, where Block’s Square ecosystem could leverage Dessa to provide hyper-personalized, low-latency customer support. However, the ‘eerily human’ nature of the model raises substantial ethical concerns regarding deepfakes and social engineering. As the industry moves toward more realistic synthetic voices, the focus will likely shift from raw capability to robust authentication and safety guardrails to prevent the weaponization of such lifelike conversational agents.

Frequently Asked Questions

Q: Can the public access the Dessa voice model right now?
A: No, Block has not released Dessa to the public. The company is currently limiting access to safety researchers to ensure the technology is tested for potential risks before a wider rollout.

Q: What makes Dessa different from other voice AI models?
A: Unlike many models that use text-to-speech pipelines or voice tags to simulate emotion, Dessa is trained entirely on human speech, allowing it to naturally convey nuance, cadence, and emotion without relying on text-based instructions.