Block's AI research lab has introduced Dessa, a research preview of a full-duplex speech-to-speech model that the company says was trained only on speech, with no text and no pretrained models. Because nothing passes through a transcript, Block says Dessa keeps the laughter, pauses, sarcasm and conversational timing that text-based voice AI drops, which could matter for voiced characters and interactive dialogue.
- Speech-only training is the core claim. Block says it trained Dessa from scratch on nothing but human conversation, and that it learned to hold a real-time conversation by listening alone.
- Block ties realism to skipping text. The company argues that converting speech to text throws away emotion and timing, and says Dessa hears and speaks those cues natively.
- Full-duplex, audio in and audio out. Block says Dessa listens and speaks at the same time and learned when to cut in and when to backchannel with responses like "yeah!" and "got it!"
- Local inference claim. Block says inference runs faster than real time and that Dessa runs locally on a MacBook Pro M4 Max or better.
- Research preview with a waitlist. Access runs through a waitlist plus early access for a small group of AI safety researchers; the page lists no API, pricing, voice cloning feature or licensing terms.
Block credits Dessa's expressive delivery to learning from speech alone, with no text step to strip out tone
Block's case is that most voice AI turns speech into text and loses information text cannot carry. The company says Dessa "natively hears and speaks emotion, sarcasm, laughter, a frustrated sigh, a voice cracking with sadness."
On the Dessa page, Block says running speech through text removes all of that, "which is why most voice AI feels stilted, slow and, worst of all, annoying." Talking with Dessa, the company says, "feels like talking with a person."
Because it learned only from speech, Block says Dessa had to pick up far more than vocabulary. Accents, pauses, breathing, humor, inflection and when to interrupt all change what a speaker means, and Block says the model had to learn them from raw sound.
Block describes speech as the harder training signal:
- Speech never repeats. The same sentence spoken twice the same way still produces different data.
- Audio is heavier than text. A single spoken word takes thousands of times more bytes than the same word written.
- Fewer words overall. Block says Dessa learned language from "hundreds of times fewer words" than LLMs need.
- One training system. The company says it trained Dessa on one NVIDIA GB200 NVL72.
Block also says Dessa knows a surprising amount, considering every fact it has came from raw sound.
Block says its unsteered demo recordings capture the timing and expression it credits to speech-only training
Every clip on its demo page is a live recording of a person talking to Dessa, according to Block, with subtitles added afterward. Block says nobody steered the model; it picked what to say, when and how, down to its own voice and personality.
Block describes these behaviors in the recordings:
- Laughs mid-sentence. Dessa laughs the moment something is funny and gets louder when asked to speak up, which Block says nobody taught it.
- Fills a memory gap. When the human speaker blanks on a movie title, Dessa finishes the sentence.
- Reacts to a surprise. Dessa gasps "oh my god" under its breath, then answers in a way Block says "sounds like teary gratitude."
- Varies intonation and speed. One clip shows a wide range in intonation and speed, and Block says a life-advice exchange "sounds genuinely empathetic."
For anyone building voiced characters or interactive dialogue, those described behaviors (mid-sentence laughter, backchanneling and a voice and personality the model picks itself) are the details worth listening for in the clips. The page does not describe any way to direct a performance or select a voice.
One demo pairs Dessa with GPT-Live-1, a third-party AI model. Block notes both voices are AI-generated and that GPT-Live-1's developer is not affiliated with Block and has not reviewed or endorsed Dessa.
Access is a waitlist and safety-researcher preview, with no API or commercial terms listed
Dessa is a research preview. The Dessa project page offers a waitlist and invites a small group of AI safety researchers to request early access to find risks the team has not anticipated.
Block lists three motivations for the project: curiosity, better understanding between humans and AI, and safety. On safety, the company argues that voice AI unable to hear emotion may keep talking while a person is in serious pain.
The lab also frames Dessa as a research subject, asking how efficiently speech-trained models learn language and facts compared to text, how efficiently they learn new languages, and what kinds of in-context learning they can perform.
Dessa comes from a small AI research lab at Block founded by former members of Dessa, the Toronto deep learning startup. The group says it has built generative speech models for over eight years and reports directly to Jack Dorsey, Block's cofounder and chairman; Ragavan Thurairatnam leads the project.
