Block, the company founded by Twitter co-founder Jack Dorsey, releases Dessa: a full-duplex conversational model trained only on human speech, which can hear emotion and runs locally in real time on a Mac
Dessa is a speech-to-speech model that Block's small team trained from scratch using only human voice: no text, no pretrained base. It is full-duplex, listening and speaking at the same time, and is claimed to run faster than real time, locally on a MacBook Pro M4 Max or better. The motivation is natural interaction, understanding emotion, and the safety risk of "continuing to chat without hearing that someone is in pain."
https://block.xyz/dessa
What it is
Positioning: reportedly the first AI to learn real-time conversation solely from human speech; the goal is to be emotional, with pauses and natural rhythm, like talking to a person.
Technical form: audio in, audio out; full-duplex (listening and speaking simultaneously); it interjects and gives backchannels (yeah / got it / ooh); the voice and personality are chosen by the model itself. The demos are live recordings of real people talking with Dessa (subtitles added afterward), with no real-time human control.
Demo highlights: laughing mid-sentence, getting louder when asked to speak up, supplying the name of a movie the other person can't remember, dropping to a whisper with "oh my god" when surprised, offering empathetic advice, and even chatting with a third-party model, GPT-Live-1 (both sides are AI voices; GPT-Live-1's developer is unrelated to Block).
Why not "speech → text → LLM"
Current voice AI often flattens speech into text, losing tone, sarcasm, sighs, sobs, and other information that text cannot carry; even the strongest models "can't hear how you feel."
The team argues this both hurts the experience and poses a safety risk: a system that keeps responding while a person is in severe pain, unaware of it.
Training and compute narrative
Trained from scratch, learning only from conversational audio; compared with LLMs, it uses hundreds of times fewer "words" (the article also notes that LLMs must learn things like C++ and Angular).
Difficulties: accents, pauses, breathing, sarcasm, and stress that changes meaning (the same sentence with different words stressed can mean entirely different things); speech almost never repeats; and a single word takes far more bytes than in text.
The entire model is claimed to have been trained on a single NVIDIA GB200 NVL72.
Why build it (Wait but why)
Curiosity: babies learn to speak without knowing how to read, so could a model do the same?
Better human-computer understanding: voice is the most natural interface, and existing approaches waste it.
Safety: as capability and autonomy grow, being nearly "deaf" to emotion is a risk; Dessa is a step toward a system that can hear emotion.