Block launches Dessa to bring natural human speech to AI

Moving beyond voice assistants

Most people are familiar with voice assistants like Siri, which often struggle with basic commands or experience long processing delays. These systems frequently feel robotic because they treat human speech as a series of commands to be converted into text. Block, the company led by Jack Dorsey, is attempting to change this dynamic with its new model, Dessa.

Dessa is designed to move closer to natural human speech by focusing on the flow of conversation. As reported by mashable.com, the model aims to bridge the gap between how humans communicate and how machines process that information. It seeks to move away from the 'flattened' experience of current voice AI.

The goal is to create an interface that feels less like a tool and more like a participant in a dialogue. By prioritizing the way people actually talk, Block hopes to make interactions with artificial intelligence feel more fluid and less mechanical.

The power of full-duplex speech

One of the primary technical distinctions of Dessa is that it is a full-duplex speech model. This means the AI can listen to a user and speak at the same time, mirroring the way two humans engage in a back-and-forth conversation. Most current systems are half-duplex, meaning they must finish listening before they can begin to respond.

To achieve this, the model was trained entirely on human speech data. This specific training method allows Dessa to mimic the natural rhythms, pauses, and timing of human interaction. It enables the model to handle interruptions and maintain a steady pace during extended discussions about complex topics like cinema.

In demonstration clips, the model shows it can provide information in a natural way rather than a robotic one. It maintains the tone of its human interlocutor while keeping up with the speed of the conversation. This ability to handle simultaneous input and output is a significant step toward realistic interaction.

Recognizing emotion and sarcasm

Dessa is also capable of identifying sophisticated vocal nuances that have historically stumped AI models. It can detect sarcasm, which requires an understanding of both the tone of voice and the relationship between that tone and the content being spoken. This is a complex task that humans perform effortlessly but is difficult for machines to replicate.

The model has shown the ability to react spontaneously to unexpected prompts. In one example, it responded to a question about an anniversary with genuine surprise and enthusiasm. It even produced human-like conversation markers, such as audible gasps and a joyful tone that bordered on tears.

The model also demonstrates a difficulty finding the right words, which is a common human trait in high-emotion moments. By capturing these small details, Dessa attempts to mirror the emotional weight of a conversation. It aims to be sensitive to the way a person sounds, not just what they are saying.

Addressing the safety problem

Block's motivation for developing Dessa involves a specific concern regarding the limitations of current voice AI. The company argues that today's technology wastes the natural interface of speech by ignoring the emotions behind the words. They suggest that current models are nearly deaf to the urgency, fear, or pain in a person's voice.

The company describes this as a safety problem. If a person is in serious pain, a voice AI that cannot detect that distress might continue its task without acknowledging the crisis. Dessa was built as a step toward fixing this by allowing the AI to recognize these emotional markers more accurately.

While recognizing a tone of urgency is not the same as feeling sympathy, it is a necessary first step for safer AI. By understanding the context of a user's voice, the AI can potentially respond more appropriately to different human needs. This advancement is intended to make AI interactions more responsible and aware.

Comparing AI speech styles

The contrast between different AI models becomes clear when comparing Dessa to other systems like ChatGPT. In a conversation between the two, the ChatGPT agent displayed robotic and unnatural speech patterns. In contrast, Dessa utilized prompts, interruptions, and puzzled questions to maintain a more human-like flow.

This comparison highlights how far Block has pushed its speech models to approximate human behavior. While ChatGPT serves as a powerful reasoning engine, its voice output can still feel distinct from a human. Dessa is specifically optimized to prioritize the nuances of the spoken word over simple text generation.

As the technology evolves, the distinction between human and machine speech may become harder to discern. Dessa represents a major advancement in bridging the gap between human communication and machine understanding. It sets a new standard for how we might interact with intelligent systems in the future.