Block unveiled Dessa, which Jack Dorsey calls "the most authentic full-duplex speech model ever built," one that "feels like talking with a person... full of emotion, nuance, pauses, and a natural cadence." Block's small AI research lab trained it from scratch on nothing but human conversation audio, with no text and no pretrained models, and says that as far as it knows, Dessa is the world's first AI with a realistic live conversation ability trained only from human speech.
What it is
Dessa is a true speech-to-speech model: audio in, audio out, with no text step in between. It is full-duplex, so it listens and talks at the same time, and having learned from real conversation, it knows when to cut in and when to backchannel ("yeah!", "got it!", "ooh"). Inference runs faster than real time, so there is no awkward wait, and it runs locally on a MacBook Pro M4 Max or better. Block says it natively hears and speaks emotion, sarcasm, laughter, a frustrated sigh and a voice cracking with sadness, all of which are lost when speech is routed through text, which is why it argues most voice AI feels stilted, slow and annoying.
The page shows live recordings of people talking to Dessa, which nobody steers: it picks what to say, when and how, down to its own voice and personality. They include:
- Laughter and volume: it laughs mid-sentence the moment something is funny, and gets louder when asked to speak up, which Block says nobody taught it.
- Recall: when a person blanks on a movie title, Dessa finishes the sentence.
- Emotion: it gasps "oh my god" under its breath at an anniversary surprise, then answers with what sounds like teary gratitude, and sounds empathetic in a conversation about life advice.
- Dessa and GPT-Live-1: in one demo it talks to OpenAI's GPT-Live-1, with both voices AI-generated. Block notes the developer of GPT-Live-1 is not affiliated with Block and has not reviewed or endorsed Dessa.
Why speech only
Block's claim is that LLMs need an insane amount of text, while Dessa learned language from hundreds of times fewer words, even though it had to learn far more than language: accents, pauses, breathing, sarcasm, humor, inflection, emotion and when to cut in. Speech never repeats, since the same sentence spoken twice comes out different, and a single spoken word takes thousands of times more bytes than the same word in text. Block says it also worked out how to train Dessa on a single NVIDIA GB200 NVL72. It gives three reasons: curiosity, since babies learn to talk without seeing a word written down; better understanding between humans and AI; and safety. AI keeps gaining power and independence while staying "nearly deaf" to emotion in our voices, and Block imagines someone in serious pain while a voice AI keeps going unaware. Researchers can now ask how efficiently such models learn language and facts compared with text, how efficiently they learn new languages and what in-context learning they can do.
Who built it
The group is a very small research lab at Block, founded by former members of Dessa, the Toronto deep learning startup. It says it has built deep learning models since 2012 and generative speech models for over eight years, and reports directly to Jack Dorsey, Block's cofounder, Chairman and "Block Head." Block is giving a small group of AI safety researchers early access to look for risks it has not anticipated. For other recent audio models, see Breeze TTS 2, and for the commercial voice race, Gemini 3.8 Live's launch charts.