AI Lab · by Salma Idrissi · 2026-10-23

Dessa: Block's conversation model trained only on speech, no text at all

Full-duplex, trained from scratch on human conversation audio - no text, no pretrained LLM. Claims it picks up emotion, pauses, natural cadence. Big open question: how efficiently does it learn facts compared to text models?

29 sparks · 4 comments

Comments

Amelia Hughes · 12 likes

would love to know the dataset licensing story here

Maya Williams · 2 likes

trained only on audio means it probably knows how to say 'umm' perfectly but not what year it is

Martina Fernández · 8 likes

honestly relatable

Amelia Hughes · 0 likes

this is the interesting direction imo. speech-to-text-to-speech pipelines throw away so much