Block, the financial technology company led by Jack Dorsey, has introduced a new artificial intelligence voice model called Dessa that aims to make conversations with AI assistants sound more natural and human.

Developed by the company’s research team, the model can reproduce expressive speech patterns, including laughter, pauses and crying.

Block said Dessa was trained entirely on human speech and can operate using a single rack of Nvidia graphics processing units, demonstrating the company’s approach to developing expressive voice technology.

Dessa uses a full duplex architecture, allowing it to listen and speak simultaneously, much like a human participant in a conversation.

Traditional voice systems often convert speech into text before processing it through a language model, potentially losing emotional cues and other information conveyed through tone and delivery.

Block argues that this approach can make interactions feel unnatural because text alone does not fully capture how people express themselves.
The company says Dessa is designed to preserve more of these conversational characteristics.

Demonstrations released by Block’s researchers show the model adopting different voices, speaking rhythms and emotional expressions depending on the conversation.

In one example, the assistant sounds tired while discussing its workload. In another, it becomes emotional while responding to a personal statement.
These demonstrations suggest that the company is attempting to build emotion directly into the model rather than relying primarily on separate instructions or labels to control how speech is delivered.

Block said Dessa can run locally on consumer computers, including devices such as the MacBook M4 Max or more powerful systems.

Local operation could reduce dependence on cloud computing, potentially lowering costs and improving privacy by allowing voice processing to take place on a user’s device.

However, the model is not yet publicly available, and the company has invited safety researchers to evaluate it, particularly for risks involving deception, impersonation and manipulation during voice interactions.

The development could have practical applications across Block’s financial technology businesses, including Square, which already uses voice AI for certain ordering tasks.

A locally operated voice model could support additional customer service and business applications while reducing the expense associated with cloud based processing.

Nevertheless, the technology also raises questions about consent, security and the potential misuse of realistic synthetic voices.

Block’s decision to seek external safety testing suggests that evaluating these risks will be an important step before the model becomes widely available.

Source: TechCrunch