Block has introduced Dessa, a speech-to-speech AI model that the company says was trained from scratch on human conversation, without text or pretrained models. Announced by Jack Dorsey on October 23, the project targets a familiar frustration with voice assistants: conversation that feels like exchanging messages, rather than talking with someone. [1][2]
The distinction is not simply a more convincing synthetic voice. Block describes Dessa as full-duplex: it can listen and speak at the same time, allowing interruptions, brief acknowledgments and other overlapping sounds that make human conversation flow. The company also says it runs locally on a MacBook Pro with an M4 Max chip or better. Those are developer claims, not independent performance results. [1]
What makes Dessa different
A text-mediated voice system converts speech into words, generates a written response and turns that response back into sound. That approach can leave out information carried by timing, emphasis, laughter or a sigh. Dessa’s stated design instead keeps speech at the center: audio goes in and audio comes out. [1]
According to Block, the model learned language and conversational behavior by listening to human speech. The team says it did not start with a pretrained language model. That makes the training approach—not just the sound of the output—the central research claim. [1]
Full-duplex interaction adds another dimension. A listener’s “yeah” or “got it” can signal attention without taking over the conversation. Block says Dessa can produce those backchannels and decide when to cut in. If that behavior works reliably outside curated examples, it could make voice interfaces less rigid and easier to use. [1]
What the demonstrations do—and do not—establish
Block’s product page presents recordings under scenarios including remembering a movie, expressive conversation and life advice. The company describes these as live recordings in which Dessa chooses what to say, when to speak and how to sound, with subtitles added afterward. It highlights laughter, changes in loudness and emotionally expressive delivery. [1]
Those descriptions explain what the team wants to demonstrate. They are not a controlled comparison showing how often the model succeeds, how it handles unfamiliar speakers or how it behaves during long, difficult conversations. Natural-sounding emotion also should not be confused with evidence of human feelings or dependable judgment.
Dorsey called Dessa “the most authentic full-duplex speech model ever built.” That is the announcement’s promotional assessment, not an independently established ranking. The more useful question is whether the same conversational responsiveness holds up under repeatable testing. [2]
The hardware claim needs context
Block says Dessa learned language from “hundreds of times fewer words” than large language models and was trained on one NVIDIA GB200 NVL72. The team itself acknowledges that text models also learn tasks beyond conversation, such as programming, which complicates a direct comparison. [1]
Crucially, one NVL72 is not one ordinary GPU. NVIDIA describes the system as a liquid-cooled, rack-scale design connecting 36 Grace CPUs and 72 Blackwell GPUs. Training on one such system is a different claim from training on a single accelerator or a laptop. [3]
The MacBook claim concerns running the model, not training it. Without matched workloads, training duration and a full accounting of compute, the announcement does not establish an apples-to-apples cost advantage over text-trained systems. Its efficiency claims are interesting research leads, not a settled economic comparison.
A conversation model, not yet proof of a complete agent
Dessa is presented by Block as a conversation model. A system that speaks fluidly is not automatically an agent that can reliably operate software, complete transactions or carry out multistep tasks. Those capabilities require their own evidence.
For businesses, the potential lies in the interface: a system that can respond to a correction without forcing the speaker to restart, or preserve vocal cues that a transcript misses. These are possible uses of the approach, not announced Dessa integrations with Square, Cash App or other Block products.
The research group describes itself as a small lab at Block, founded by former members of the Toronto deep-learning startup Dessa and reporting directly to Dorsey. The announcement is signed by Dessa lead Ragavan Thurairatnam. [1]
Why safety testing matters
Block frames better awareness of vocal emotion as a safety opportunity, using the example of an AI failing to notice that someone is in pain. It is inviting a small group of AI safety researchers to seek early access and investigate risks the team has not anticipated. The site also offers a waitlist. [1]
That invitation is important because expressive speech can make a system feel more trustworthy without making its answers more accurate. Evaluation should examine factual reliability, responses to distress, interruption handling and the risk that users mistake an expressive simulation for understanding.
Dessa’s most consequential idea is therefore narrower—and more interesting—than a claim that AI now talks exactly like a person. Block is testing whether learning directly from speech can produce a different kind of conversational interface. The next milestone is evidence that its natural timing and expressiveness translate into reliable, useful behavior beyond the showcase recordings.
Comments (0)