The Hume Audio Decision API is a new audio-native model that answers arbitrary questions about speech and sound—such as speaker count, laughter, frustration, tone, or custom perceptual criteria—in a single API call, returning calibrated probabilities. Trained on 10M+ human judgments, it supports realtime streaming or batch processing of long recordings and is designed for voice AI evaluation, contact centers, and agentic systems. Access is currently preview-only via request/waitlist.
Hume’s Audio Decision API accepts an audio recording and developer-defined questions, then returns structured probabilities rather than a transcript or free-form explanation. Questions can cover speech, voices, emotion, background sounds, conversation dynamics, and other audio judgments; supported formats include yes/no, multiple choice, scales, and event time spans. The documentation describes uses such as evaluating text-to-speech, testing voice agents, reviewing support calls, and labeling speech datasets. Several questions can be asked about one recording in a request, and the results can be used in code to rank, threshold, or combine decisions.
