Block’s Dessa AI voice model learned to talk without text: Here’s how
Block's Dessa is speech-to-speech model trained only on human conversation, with no text or pretrained models
Dessa learnt to talk without ever reading a word. Block, the company behind Square and Cash App, says its new AI model was trained from scratch on nothing but human conversation, with no text and no pretrained models.
The result, Block claims, is a voice that laughs mid-sentence, gasps and cuts in when it should. The Block Dessa AI voice model is a research preview, and every demo so far comes from Block itself.
How the Block Dessa AI voice model works?
Voice assistants normally convert your speech to text, create a response and speak it out. Dessa does not use text at all. The system accepts audio input and gives audio output.
According to Block, in this way it keeps something that text fails to preserve: sarcasm, humour, an exasperated sigh, and a broken voice from sadness. Moreover, Dessa operates fully duplex. It listens and speaks simultaneously; hence, it can add such words as “Yeah” or “Got it” without waiting for a pause.
According to Block, Dessa mastered language with thousands of times fewer words than the language models do. The training was done on a single Nvidia GB200 NVL72 system.
Speech is a harder teacher than text, Block argues. A spoken word takes thousands of times more bytes than the same word typed, and a sentence never sounds exactly the same twice.
In one clip, Dessa finishes a person's sentence when she blanks on a movie title. In another, it answers an anniversary surprise with what sounds like teary gratitude.
Team lead Ragavan Thurairatnam also posted a block intern playing a guessing game with the model. Dessa guessed her object on the first try, twice, then kept guessing after being told it had won.
One demo pits Dessa against GPT-Live-1, a third-party model. Block says that model's developer hasn't reviewed or endorsed Dessa.
Can you run Dessa locally?
Block says Dessa runs on a MacBook Pro M4 Max or better. A small group of AI safety researchers has early access.
The company's safety argument is blunt. Today's AI is nearly deaf to the emotion in your voice, so a voice assistant could keep talking while you're in serious pain.
