Hume AI's Audio Decision API lets developers send raw audio and ask plain-language questions about speakers, laughter, frustration, tone, and background sound. It is aimed at teams building voice agents, scoring contact-center calls, or evaluating generated speech. One request can cover many questions at once and returns a probability for each answer the developer defines, as yes or no, a choice, a rating, or a time span, so the result can drive a rule or an agent without a transcript to interpret. Voice products have mostly worked from words, which leaves out the pause, the sigh, the overlap, and the tone that often explain why a call went badly. Collecting those signals used to mean running a stack of separate models and combining their outputs. Hume's model takes raw audio and answers in a single call, in real time or across hours of recordings. It is trained on more than 10 million human judgments of how people perceive speech and sound. Questions can be added or changed without retraining, so a contact center can rescore an archive against a new rubric and a live agent can decide whether to wait, clarify, or escalate from the audio it already has. Hume AI, based in New York, builds voice models and measurement tools from emotion research and large panels of human ratings on expression, naturalness, and delivery. The Decision API puts that work into an application as a structured judgment with a confidence score that code can threshold.