Block has introduced Dessa, a full-duplex AI voice model that mimics human speech, emotions, and laughter while running locally on consumer hardware. The technology aims to reduce reliance on cloud-based infrastructure for automated customer interactions. While the company has not released it publicly, the model is undergoing safety testing to address potential risks like voice-based fraud and social engineering.
Block, Inc. has unveiled a new AI technology called Dessa, which is designed to handle speech-to-speech interactions with a focus on human-like realism. Unlike many current voice AI systems that convert text into speech, often resulting in a robotic tone, Dessa is trained directly on human audio. This allows it to naturally express pauses, laughter, and varying emotional states, which the company claims makes the interaction feel more authentic.
A key aspect of this development is the model’s ability to operate locally on high-end consumer hardware, such as the MacBook M4 Max. This marks a potential shift in how companies deploy large AI models. By running these processes on the device itself rather than relying entirely on massive, cloud-based computing, the company could reduce infrastructure costs and latency for voice interactions. For investors and market observers, the ability to lower the cost of operations while improving automated services is often a key area of interest, especially for a financial services company with a platform as large as Square.
The system utilizes full-duplex technology, which allows it to listen and speak at the same time. This mimics natural human conversation, where overlaps and interruptions are common. The company envisions that this technology could be integrated into its financial ecosystem, potentially assisting with automated customer support for the small businesses that rely on the Square platform for commerce.
Despite the technical capabilities, Block has chosen not to release the model for public use yet. The company is instead inviting safety researchers to test the technology. This cautious approach is driven by the significant risks associated with hyper-realistic AI voice agents, specifically the threat of fraud, impersonation, and social engineering. As governments and regulators worldwide increase their scrutiny of generative AI, the ability to control and secure such tools is a major factor for any company entering this space.
Moving forward, investors will likely track whether the company can successfully commercialize Dessa while maintaining strict safety standards. The core monitorables include the timeline for potential deployment, the actual reduction in computing costs compared to traditional cloud-based voice assistants, and how the company manages the regulatory challenges surrounding AI-generated speech. The success of this initiative will depend on balancing the product's functional benefits for merchants with the necessity of preventing malicious use.