Jack Dorsey’s Block is previewing a voice AI model that could change how businesses build conversational interfaces, though its reliability remains to be demonstrated beyond company-selected examples.
Block has introduced Dessa, an experimental AI voice model trained entirely on recordings of human conversations, without written text or pretrained models. The company says the system learned to communicate by listening to people talk rather than relying on the text-based training methods used by conventional language models.
According to Block, Dessa is a full-duplex speech-to-speech model, meaning it can listen and speak simultaneously. Block says it processes audio directly instead of converting speech into text before generating a response, retaining vocal information such as timing, laughter, and intonation that a transcript may omit.
Block says Dessa can react to jokes mid-sentence, adjust its volume when asked, and respond with different speaking styles depending on the conversation. The company also says it trained the model using one NVIDIA GB200 NVL72 rack-scale system containing 72 GPUs and that it can run locally on a MacBook Pro equipped with an M4 Max chip or better.
Why Block is rethinking voice AI
A common design for voice assistants converts speech into text, generates a written response, and converts that response back into speech. Block argues that this approach loses information conveyed through tone, timing, and other vocal signals.
“We trained Dessa from scratch on nothing but human conversation. Sounds! No text, no pretrained models. It learned to hold a real-time conversation simply by listening,” the company said in its announcement.
Block is also exploring whether speech-only training could offer a different path to teaching AI language and conversational skills. The company says Dessa learned language from hundreds of times fewer words than large language models typically require, although its announcement does not provide benchmark results or a detailed comparison methodology.
Natural conversations bring new safety concerns
Block’s demonstrations show Dessa responding with laughter, pauses, and emotional inflections. However, these company-selected examples do not establish how consistently the model handles real-world conversations.
That distinction matters because sounding human does not necessarily mean understanding a person’s emotional state. A model that convincingly imitates distress, sympathy, or excitement could make it harder for users to distinguish synthetic speech from genuine human communication.
Another concern is misuse: increasingly realistic AI voices could make scam calls more convincing. However, producing natural speech does not itself demonstrate an ability to clone a specific person’s voice.
Block says it will give a small group of AI safety researchers early access to investigate risks before wider access.
More must-read AI coverage
- SS&C Intralinks DealCentre AI vs. Datasite: Which platform is built for the future of dealmaking?
- SS&C Intralinks FundCentre AI vs. Juniper Square: Which platform better supports modern private markets fund managers?
- Why Data, Not Models, Determines AI Success
- The Rise of the AI-Native Factory: How Physical AI Is Transforming Manufacturing
What Dessa could mean for users and businesses
For businesses evaluating natural voice AI, smoother turn-taking could make automated support and ordering less frustrating. The practical test is whether a system captures requests correctly, handles interruptions, and recognizes when a caller needs human help.
Block operates Square, which already offers AI-powered voice ordering. That provides relevant business context, but does not establish that Dessa powers the service or that an integration is planned.
Local execution could allow some speech processing to happen on a business-controlled device. However, it would not automatically make a complete service offline or private: connected ordering systems, customer databases, logging tools, and payment services could still transmit or retain information.
Before considering a future deployment, IT teams should seek tests covering background noise, varied accents, incorrect answers, and escalation to human staff. They should also assess response times under sustained use and the total cost of hardware, integration, monitoring, and support.
For now, Dessa is a research development rather than a purchasing option. Block’s announcement does not provide commercial pricing or a general release date. Businesses should watch for independent evaluations and deployment documentation before treating its conversational demonstrations as evidence of production readiness.
Read more: Explore why AI guardrails need data boundaries and how scoped permissions and audit logs can help limit unauthorized access.