Block’s Dessa voice model talks and listens at the same time

Most voice assistants transcribe what you say, have a language model write a reply, then read it out. Dessa works on the sound itself, so pauses, emotion and sarcasm that a transcript loses stay in, The Decoder explains. It is also full duplex: it listens while it speaks, the way people cut in on each other. Block trained it only on recordings of human speech, on a single rack of Nvidia GPUs, and it runs on a MacBook Pro with an M4 Max chip or newer. “The most powerful AIs in the world can’t hear how you feel,” Block wrote.

There is no public access. Block has asked safety researchers to try the model first, TechCrunch notes. TechCrunch sets it against ElevenLabs v4, where written tags in the script set the emotion, and says it is not clear whether Dessa’s emotion is steered some other way.

For ad voice-overs, the paid tools on our list of AI voice-over generators start low: ElevenLabs Starter costs $6 a month for 30,000 credits and includes a commercial licence, read on 23 October 2026.

Until Block opens Dessa beyond safety testers, emotion in an ad voice-over still comes from tag-driven tools such as ElevenLabs, whose Starter plan costs $6 a month with a commercial licence.