Hume Audio Decision API Turns Audio Into Decisions

The breakthrough is making subjective listening programmable: developers can ask whether someone sounds frustrated, whether speakers overlap, or how natural a voice sounds, then route on the resulting probabilities.

• –Native audio preserves tone, pauses, laughter, overlap, and background sound that transcript-first pipelines discard.
• –Typed outputs such as yes/no, choice, score, and time span are directly composable in application logic.
• –One request can answer multiple questions about the same recording, reducing brittle stacks of task-specific classifiers.
• –Hume claims 70.7% agreement with human panels versus 56.5% for individual raters, but teams should still validate thresholds on their own domains.
• –The waitlist-only launch positions this first as voice-agent evaluation and operations infrastructure, not a general conversational model.