Block’s Dessa aims to make local voice AI feel remarkably human

Introduction

Voice AI is moving beyond systems that simply convert text into spoken words. Block’s research team has released an early look at Dessa, a model intended to conduct conversations directly in speech. In demonstrations shared by the company and its researchers, the system uses laughter, pauses, changes in cadence, tiredness, and highly emotional delivery. The result is designed to sound less like a voice assistant reading a script and more like a person responding in real time.

Key points

• Full-duplex conversation: Dessa can listen and speak concurrently. That allows it to respond more like a human conversational partner rather than waiting for a complete turn before processing the next one.
• Speech-to-speech design: Block says the model was trained entirely on human speech. The company also says its training used a single NVIDIA GPU rack, although it has not disclosed the model’s size or full training configuration.
• Emotion without a conventional text middle layer: Many voice systems first generate text and then use that text, sometimes together with emotion tags, to produce audio. Block argues that this process can discard information carried by pitch, timing, hesitation, and vocal tone. Dessa is intended to keep more of that information in the speech interaction itself.
• Local inference: Block says the model can run on a consumer computer such as a MacBook M4 Max or better. It has not yet published detailed benchmarks covering latency, memory use, or the exact hardware requirements.
• No public release yet: Dessa is currently unavailable to the general public. Block is inviting safety researchers to test how a voice agent might mislead people or be mistaken for a human.

Why it matters

Local execution could be one of Dessa’s most important practical features. If a capable voice model can operate on a personal computer, applications may be able to reduce cloud dependence, network delay, and inference costs. Local processing could also make it easier to place voice interaction inside business tools and consumer devices, provided that developers can manage the model’s hardware demands and data handling. Block’s Square business already uses voice AI for ordering, giving the company a direct commercial context in which this technology could be relevant.

However, a convincing voice is not only a usability improvement. It can also change how people judge an automated system. A model that sounds tired, vulnerable, or emotionally affected may encourage users to attribute more awareness or trustworthiness to it than is warranted. That makes identity disclosure, consent for recording, abuse monitoring, and safeguards against impersonation important parts of the product design.

Dessa should therefore be viewed as a technology demonstration rather than a mature general-purpose assistant. The available material does not establish how well it handles factual questions, long conversations, interruptions, or difficult safety cases. It is also unclear how much of its apparent emotion is intrinsic to the model and how much is controlled by undisclosed mechanisms. The next test will be whether Block can preserve natural interaction while making the system transparent, controllable, and safe to use.

Source: TechCrunch AI