The image generated by AIYXL using AI
A pause can signal hesitation. A quick interruption can show recognition. Written transcripts preserve the words while leaving much of that interaction behind.
Block’s Dessa experiment starts with the sound itself. The company says it trained the model from scratch on human conversations, using neither text nor pretrained models.
Dessa processes incoming and outgoing audio and can listen while speaking—a capability known as full-duplex conversation. Block’s recordings illustrate changes in delivery, brief interjections and overlapping speech. The company says the model can run locally on a MacBook Pro M4 Max or better.
The research question reaches beyond producing a convincing voice: how much language, knowledge and conversational behavior can a model acquire through listening?
For now, the evidence consists of demonstrations selected by Block. The launch page provides a waitlist and invites a small group of AI safety researchers to investigate unexpected risks.
Expressive speech can make a system compelling. Whether it understands a speaker reliably across unfamiliar situations requires broader testing.
Sources: Block