Dessa: Block unveils a voice model driven only on speech, which runs on a MacBook

Block's IA lab presents Dessa, a full-duplex speech-to-word model learned without any text, able to listen and respond at the same time on a laptop.

Block's internal research laboratory, Jack Dorsey's company, presented a voice conversation model on October 23, Dessa, which receives audio and returns audio without a written transcript. His particularity: he was trained from zero on recordings of human conversations, without textual corpus or pre-entered model at the starting point. The team, led by Ragavan Thurairatnam and from the Toronto start-up Dessa bought by Block, talks about a "GPT-3 moment" for speech.

Listen and talk at the same time

Dessa works full-duplex: he can speak while listening, and would have learned real exchanges when to cut the word or drag a "hm hm" to show that he follows. Block claims that he perceives and restores emotions, irony, laughter or a broken voice. The training took place on a single Nvidia GB200 NVL72 system, and the inference, faster than real time, runs locally on a MacBook Pro M4 Max or more powerful.

The model is not yet open to the public: a small group of AI security researchers have access to it as a preview and a waiting list is open. The first demonstrations show a very natural look, but also some awkward exchanges during the subject changes.