The voice that refuses to wait its turn #
Block, Inc., Inc., the company behind Square and Cash App, introduced Dessa, a full-duplex conversational speech model that redefines human-AI interaction. This system can listen and speak simultaneously, enabling a fluidity previously absent in AI conversations.
Unlike traditional voice assistants that operate like a walkie-talkie—listening, processing, then speaking—Dessa breaks this rigid turn-taking. It facilitates natural human conversational elements: backchannels like "yeah" and "got it," spontaneous interruptions, and even laughter mid-sentence.
Demos showcase Dessa’s startling conversational fluency, but this is a testament to its advanced processing, not human-level intelligence. The model, for instance, can still make errors, such as confusing a speaker with someone else, as seen in a demo where it mistakenly identified its interlocutor as "Ronaldo."
This full-duplex capability, trained entirely on raw audio from human conversations, marks a significant departure from text-based models. Dessa directly processes sound in and out, bypassing the text bottleneck that strips away emotional nuance and prosody, resulting in an unnervingly human-like conversational experience.
Why skipping text changes the sound #
Conventional AI voice assistants operate like a relay race: speech recognition converts audio to text, a language model processes the text, and text-to-speech then synthesizes a reply. This cascaded pipeline discards crucial elements of natural conversation. Tone, s, emotion, volume shifts, laughter, and the subtle timing of interjections all vanish when speech reduces to text.
Dessa sidesteps this text bottleneck entirely. It processes audio directly, taking sound in and sending sound out without an intermediate text layer. This direct approach allows it to retain and replicate the full richness of human vocal nuance, including the "yeahs" and "got its" that punctuate real-time dialogue.
Block, Inc., Inc. claims Dessa was trained from scratch on recordings of human conversations, bypassing text datasets or pre-existing language models. This audio-first training allowed the model to learn conversational dynamics, such as when to laugh mid-sentence or interrupt naturally, directly from acoustic patterns.
Remarkably, Block, Inc. reports Dessa can perform local inference faster than real-time on consumer-grade hardware, specifically an Apple MacBook Pro with an M4 Max chip. This deployment claim suggests a highly efficient architecture, a stark contrast to the large-scale GPU clusters typically required for cutting-edge generative AI models.
The uncanny demo has a safety catch #
Conversational realism makes an AI feel more trustworthy than a polished, obviously synthetic voice. Dessa's natural rhythm and human-like interjections, like saying "yeah" or laughing mid-sentence, can make it sound genuinely engaged. This mimics human interaction, creating a strong illusion of understanding, even when the underlying AI lacks true comprehension or simply fabricates information.
A Block, Inc., Inc. demo showcased this risk. Dessa, when discussing its own capabilities, invented a human-sounding backstory about its "impressive" expressiveness and ability to "understand emotions much better than anything I've seen out there." The natural delivery lent this fabricated self-assessment an undeserved credibility, highlighting how effortlessly an AI can present generated content as fact.
This blend of voice cloning and responsive dialogue poses a significant threat. Scammers could leverage Dessa-like technology to create hyper-realistic, interactive voice calls for family-emergency, bank, delivery, or tech-support scams. The naturalistic speech patterns, including s and interruptions, could bypass the typical "robot voice" red flags, especially for individuals unfamiliar with current AI capabilities. For further technical details, explore Dessa - The Conversation Model Trained Only on Speech.
Imagine a scammer cloning a grandchild's voice, then having a full-duplex AI engage a grandparent in real-time, complete with emotional cues and natural conversational flow. The ability to interrupt, respond to emotional shifts, and even laugh could make these malicious interactions virtually indistinguishable from genuine human contact, dramatically increasing their effectiveness.
Enjoying this? Get one like it in your inbox each morning.
one email a day · unsubscribe in two clicks · no third-party tracking
A preview today, a trust problem tomorrow #
Block, Inc., Inc. has released Dessa as a controlled research preview, not as public model weights or a general API. Access is limited to safety researchers, reflecting the inherent risks of such a powerful tool. This cautious approach acknowledges the technology's potential for misuse.
While safeguards like detection and audio watermarking are under development, they face an ongoing arms race. The challenge lies in creating robust, ubiquitous defenses against rapidly evolving synthetic audio. No single solution can fully mitigate the risks. Consumers must adopt proactive safety measures. Always independently verify urgent requests by calling a known number, never relying solely on a caller's voice. Establishing a family code word for critical communications provides an additional layer of protection.
The AI-to-AI demo, where Dessa converses with another AI, hints at applications beyond scams, such as advanced virtual assistants or synthetic media creation. This suggests a future where AI voices become integral to various digital interactions, necessitating constant vigilance and adaptation.
Frequently Asked Questions #
What is Block’s Dessa?
Dessa is a conversational speech model from Block designed to listen and speak at the same time, with natural timing and vocal cues.
How is Dessa different from a typical voice assistant?
Block says Dessa was trained on audio conversations and generates speech directly, rather than relying on a speech-to-text, language-model, and text-to-speech relay.
Can the public use Dessa now?
According to the supplied reporting, Dessa is in a limited research preview, with early access focused on AI safety researchers.
Could realistic AI voices be used for scams?
Yes. A convincing, responsive voice could make impersonation and social-engineering calls more persuasive, so callers should verify urgent requests through a separate trusted channel.