Block unveils Dessa, a voice AI that learned from conversations alone, with no text

What this is about

Block, the company of Twitter co-founder Jack Dorsey (and operator of Square, among other things), has unveiled Dessa, an experimental speech model. According to the company, it was trained entirely on recordings of human conversations, with no written text and no pretrained models. For now, Dessa can only be seen in demos. You can't use it yourself.

"We trained Dessa from scratch on nothing but human conversation. Sounds! No text, no pretrained models. It learned to hold a real-time conversation just by listening."

How Dessa works

Typical voice assistants work in three steps: speech is converted to text, a language model writes a text reply, and that reply is read aloud. Block argues that a lot gets lost along the way: emphasis, laughter, breathing, pauses, sarcasm, mood. Dessa, by contrast, is a speech-to-speech model: it takes in audio and produces audio directly.

It is also full-duplex. That means it can listen and talk at the same time, like people on a phone call. According to Block, this lets it sense when to jump in, drop in backchannel cues like "yeah!" or "got it!", laugh mid-sentence when something is funny, adjust its volume on request, and sound different depending on the conversation. Block stresses that the emotional expression is built into the model and is not steered with control tags (text markers for emotions). Whether something similar is used internally is unclear, according to IBTimes.

Training and hardware

Block says Dessa learned speech from hundreds of times fewer words than large text models typically need. The announcement gives no benchmark or comparison method. Training reportedly ran on a single NVIDIA GB200 NVL72, a rack with 72 GPUs. Inference is said to work locally on a MacBook Pro with an M4 Max or better, with no cloud. The team justifies the project with an analogy:

"Babies learn to speak without ever seeing a written word. We wanted to know if a model could too. It can. It's just very hard."

The demos

Block shows examples with different voices, speaking styles and moods. In one conversation, Dessa says it's tired from work, and it sounds like it. One example shows a kind of emotional outburst just before crying. Other clips show conversations about movies, guessing games, and a conversation with OpenAI's voice model GPT Live 1, presented by project lead Ragavan Thurairatnam. According to Cryptobriefing, some demos also show weaknesses: the model mixes up speakers or abruptly changes the subject.

Team, safety and business angle

The small research team at Block is led by Thurairatnam. Members previously worked at the Toronto AI startup Dessa and report directly to Dorsey. Block is first giving a small group of AI safety researchers access to look for unexpected risks. According to TechRepublic, possible problems include these: a model that convincingly imitates distress or empathy makes it harder to tell real human communication apart, and more realistic voices could make scam calls more believable. However, producing natural speech doesn't prove that specific people can be cloned.

For Block, a locally running voice model is also commercially interesting: Square already offers AI voice ordering, and running locally could save expensive cloud inference. Whether Dessa will be used there is not documented, according to TechRepublic. Also, locally run speech processing isn't automatically offline or private if it is connected to ordering, customer and payment systems.

Our take

Dessa is technically intriguing because it suggests conversational ability may emerge even without text pretraining. But so far that is a vendor claim based on selected demos. Independent tests, benchmarks, weights, pricing and a release date are all missing. For businesses, this is currently a research development, not something you can buy. Wait for tests with background noise, accents, wrong answers and handoff to humans. Response times under continuous operation and total costs are also still open questions.