Block releases Dessa, a conversation model trained only on speech
Block published a post announcing Dessa, a conversation model trained only on speech. The source is Block's own page at block.xyz/dessa, and the post was discussed on Hacker News with 362 points and 134 comments.

Why it matters
A speech-only training approach is a distinct design choice from text-first conversation models, which may affect voice agent and spoken interface work. Builders evaluating voice features can track how this model compares on their own tasks.
Who should care
Voice agent developers, call center and customer support tooling teams, product teams building spoken interfaces
What you can do
Read the Block Dessa post and note its stated training data, capabilities, and limitations, then test it against your current voice pipeline on a sample of real user utterances.

