Updated
Updated · The Verge · Oct 23
Block Unveils Dessa Voice Model Trained on 100% Human Conversation
Updated
Updated · The Verge · Oct 23

Block Unveils Dessa Voice Model Trained on 100% Human Conversation

3 articles · Updated · The Verge · Oct 23

Summary

  • Block introduced Dessa, a speech-to-speech AI model built from scratch using only human conversation, and released demo chats showing natural but sometimes awkward exchanges.
  • Dessa runs locally on a MacBook Pro M4 Max or better, with Block positioning it as a voice-native system that avoids flattening speech into text.
  • The company said it built the model for curiosity, better human-AI understanding and safety, arguing current voice AI often misses timing, emotion and signs of distress.
  • Block plans to give early access to a small group of AI safety researchers as it tests whether conversation-only training can make voice AI feel more human and more responsive.

Insights

How does a textless AI like Block's Dessa verify factual accuracy when learning everything purely from raw human audio?
If Dessa bypasses text to preserve human emotion, does it become an un-auditable black box for AI safety researchers?
Can an AI model trained solely on human speech truly understand language without the visual or textual context humans rely on?