Dessa: Speech to Speech, No Text in Between
Block just released Dessa, a speech-to-speech AI model trained on nothing but human conversation—no text, no pretrained language model, no transcript in the middle. For creators who need voices that sound human, respond in real time, and carry emotion, it's a genuine shift.
What Dessa actually does
Block's in-house AI research lab announced Dessa on October 23, 2026, as a full-duplex speech-to-speech model trained from scratch on nothing but human conversation. The key difference: it's true speech-to-speech with no speech-to-text or text-to-speech step in the middle.
The model uses a full duplex architecture, meaning it can listen and speak simultaneously, and can recognize laughter, sarcasm, changes in tone and emotional expressions that can be lost when speech is converted into text.
The demos are the evidence. Dessa laughs mid-sentence the way people do; asks to speak up and gets louder without being prompted; finishes incomplete thoughts; and in an anniversary clip gasps "oh my god" under its breath and responds with what sounds like teary gratitude. It drops in backchannels like "yeah!", "got it!" and "ooh" while you're still talking, to show it's following.
What creators see here
For voiceover work, character dialogue, and interactive projects, Dessa solves a real friction. Current voice AI flattens what people say into text and discards information that text alone cannot carry—current systems strip out emotion, sarcasm, laughter, pauses, and the timing that gives speech its meaning. Dessa preserves all of it.
One demo has Dessa chatting with GPT-Live-1, a third-party voice model—both AI voices, real conversation. Another shows it handling movie discussions, guessing games, and personal stories with timing and interrupts that feel natural.
The efficiency is notable too. The model learned language using hundreds of times fewer words than conventional text-based language models, and was trained using a single NVIDIA GB200 NVL72 system.
What's actually good, and what isn't
Local inference is real. Dessa can run locally on a MacBook Pro equipped with Apple's M4 Max chip or better, without requiring cloud infrastructure for inference. It listens and talks at the same time, cuts in and backchannels naturally, runs faster than real time. No latency, no cloud, no data leaving your machine.
But: this is a research preview, not a product. There is no public API or weights yet; a small group of AI safety researchers is getting early access. The launch page does not pair that claim with a public installation guide, memory requirement or downloadable model.
The demos are also curated. The demo conversations sometimes feel like talking to someone still learning social cues—Dessa can maintain topic coherence and respond appropriately to direct questions, but its responses occasionally miss subtle contextual signals that human speakers navigate instinctively. One conversation mistakenly addressed its interlocutor as Ronaldo at the beginning.
What the demos cannot establish is how often it behaves naturally when nobody is selecting the clip.
How to try it
Early access is limited to a small group of AI safety researchers, who must apply with a name and an institutional email from a university, company or lab. Block says it is giving that group access "to find the risks we haven't anticipated." Broader access remains a waitlist.
You can join the Block Dessa waitlist and watch the demo conversations to get a sense of how it sounds. Block said an app is coming soon.