Blocks latest AI model Dessa was trained entirely on human speech - and it shows

Block just dropped Dessa, an AI model trained exclusively on human speech. This approach moves away from standard text-heavy training to prioritize natural, audio-first interaction.

If voice-first AI becomes the standard for fintech or customer ops, Dessa could solve the uncanny valley problem. Builders need to watch if this lowers the barrier for high-quality voice agents.

This move puts pressure on OpenAI and ElevenLabs to accelerate low-latency, speech-native reasoning. It shifts the competition from pure model intelligence to how naturally an AI can exist in audio environments.

• →Develop voice-first customer support agents that bypass the robotic, scripted feel of current LLM-to-speech workflows.
• →Build niche audio-to-action tools for hands-free industries like logistics or healthcare where typing is a bottleneck.
• →Integrate speech-native reasoning into existing fintech apps to simplify complex transactions through natural conversation.

• →Data moat scarcity, as competitors can replicate speech-only training once they secure similar audio rights.
• →Integration friction, specifically the risk that users might distrust voice-only confirmation for high-stakes financial tasks.