Block releases Dessa, a full-duplex model trained purely on speech, which can run locally in real time on a MacBook Pro M4 Max
Block, founded by Twitter co-founder Jack, has released Dessa, a speech-to-speech model trained from scratch purely on human speech, with no text and no pretrained base, supporting full duplex so it can listen and speak at the same time. The model claims inference faster than real time and can run locally on a MacBook Pro M4 Max and above, and the entire model is said to have been trained on a single NVIDIA GB200 NVL72. The team says its goal is to let the model understand tone and emotion, avoiding the safety risk of continuing to respond without perceiving the other person's distress.
Why read this
AI score 76
The original text sets out three points—training purely on speech, full duplex, and local real-time operation—from which one can infer the approach of a speech model that bypasses the text intermediate layer.
Topics
View original postWe believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence. https://t.co/r0qFLdvd6c
View original
We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence. https://t.co/r0qFLdvd6c
Body · AI translation
Block, founded by Twitter co-founder Jack, releases Dessa: a full-duplex conversational model trained purely on voice, able to hear emotion, running locally in real time on a Mac
Dessa is a speech-to-speech model trained from scratch by a small team at Block using only human speech: no text, no pretrained base, full-duplex listening and speaking at the same time, claimed to infer faster than real time, and able to run locally on MacBook Pro M4 Max and above; the motivation is natural interaction, understanding emotion, and the safety risk of "continuing to chat without hearing the pain." https://t.co/sOZgeLlnTa
# What it is
Positioning: currently claimed to be the first AI that learns real real-time conversation solely from human speech; the interaction goal is to carry emotion, pauses, and natural rhythm, like talking to a person.
Technical form: audio in, audio out; full-duplex (listening and speaking simultaneously); it interjects and backchannels (yeah / got it / ooh); voice and personality are chosen by the model itself, and the demo is a live recording of a real person with Dessa (added after the subtitles), with no one controlling it in real time.
Demo highlights: laughing mid-sentence, getting louder when asked to speak up, picking up on a movie name the other person couldn't recall, lowering its voice with an "oh my god" when surprised, empathetic advice, and even chatting with a third-party GPT-Live-1 (both are AI voices; the developer of GPT-Live-1 is unrelated to Block).
# Why not do "speech→text→LLM"
Current voice AI often compresses sound into text, losing information that text cannot carry, such as tone, sarcasm, sighs, and choking up; even the strongest models "can't hear how you feel."
The team believes this both harms the experience and poses a safety hazard: when a person is in severe pain, the system notices nothing and keeps responding.
# Training and compute narrative
From scratch, learning only from conversational sound; compared with LLMs, it uses hundreds of times fewer "words" (the text also notes: LLMs also have to learn C++, Angular, etc.).
Difficulties: accents, pauses, breathing, sarcasm, and stress changing meaning (the same sentence with different words stressed means something completely different); speech almost never repeats; the byte count per word is far higher than for text.
Claims the entire model was trained on a single NVIDIA GB200 NVL72.
# Why do it (Wait but why)
Curiosity: babies can learn to speak without knowing how to read; can a model do the same?
Better human-machine understanding: voice is the most natural interface, and existing approaches waste it.
Safety: as capability and autonomy rise, being almost "deaf" to emotion is a risk; Dessa is a step toward "hearing emotion."