Author: SRAM
Robotic phone calls used to be easy to spot. A stiff voice, awkward pauses, and a script that continued before you had finished speaking were often enough to make you hang up.
But when AI starts pausing naturally, picking up conversational cues and responding with emotion in its voice, can you still judge who is on the other end simply by how human they sound?
Block, the company led by Jack Dorsey, recently introduced the voice AI model Dessa. According to its official description, it was trained from scratch on human conversation audio, without text or pretrained models, with the aim of preserving tone, emotion and natural rhythm. It supports full-duplex interaction: listening while speaking, rather than always waiting for one side to finish. It remains a research preview.
What makes this direction interesting is how much meaning a voice can carry.
The same words—“no problem”—can convey enthusiastic agreement or reluctant acceptance. “I’m fine” can reassure you, or make you more concerned if the speaker’s voice is trembling.
Reduce a conversation to text and those differences can disappear. Making voice AI feel natural involves much more than clear pronunciation.
The potential for everyday use is appealing. Imagine calling to change an appointment and correcting yourself halfway through, with the system following the change. Or having it ask a timely question when your request is unclear, instead of making you start again. These are experiences worth pursuing, although they still need to be tested in real applications.
But as a voice becomes better at putting us at ease, trust may form too quickly.
A voice that sounds patient and empathetic does not make its answers correct. Nor does it establish that the speaker is authorised to represent a bank, a company or someone you know.
The issue does not even require impersonating a particular person’s voice. A fluent conversation with natural responses may be enough for us to turn “it seems to understand me” into “I can trust it.”
From a product-design perspective, identifying an AI clearly becomes more important as voice technology improves. Businesses considering it for customer service need to think beyond a friendly voice: when should a person take over? How should payments or changes to records be confirmed? Who is responsible if the system says something wrong?
The direction demonstrated by Dessa makes talking to computers more promising. But a natural voice should never replace identity checks and sound judgment.
When AI speaks to you in a reassuring voice, are you trusting what it says—or simply the way it says it?
Source: Block | Dessa research preview. This article includes independent editorial analysis by SRAM.