Daily AI CatchupBlock's Dessa Learns Speech From Sound Alone
Block built Dessa, a conversational AI trained entirely from scratch on human speech audio, with no text used anywhere in training. The model learned to hold real-time conversations, laugh on cue, follow instructions, and complete sentences, picking up facts purely from what it heard in recordings.
Why it matters
💻 Developer · A rare example of a large model trained purely on audio, useful if you're exploring speech-native architectures instead of text-to-speech bolted onto an LLM.
📦 Product · Voice assistants built this way could sound more naturally conversational than today's text-pipeline voice bots.
🎨 Design · Worth listening to the example conversations, this could change expectations for how natural voice interfaces should feel.
📈 Business · Block entering voice AI signals more competition in a space already crowded by ElevenLabs and Mistral.
🤔 Just Curious · An AI learned to talk, laugh, and hold conversations just by listening, never reading a single word.
Sources: Dessa - the Conversation Model trained only on speech