Hume Turns Raw Audio Into Decisions: New API Answers Questions About a Call Without Transcribing a Word

Hume AI, the startup built on years of emotion-perception research, launched an Audio Decision API on November 2, and the pitch is simple: stop transcribing. Developers send raw audio along with a question and a set of possible answers, and the model returns probabilities over those answers. Does the caller sound frustrated? Is there a siren in the background? Where exactly does the laugh start and end? No transcript, no labels, no retraining.

The company calls it the first audio-native model that answers questions about speech and sound. Questions can be yes or no, multiple choice, rating scales, or time spans, and adding one is a line of JSON, so a team can change what it measures on a live call in seconds instead of rebuilding a rubric pipeline. Confidence thresholds can be wired straight into business rules: wait, clarify, or escalate.

On accuracy, Hume reports that the Audio Decision model agreed with a human panel average judgment 70.7 percent of the time in an initial evaluation, against a 56.5 percent baseline for agreement with a single individual rater. That evaluation is Hume's own, built on the same panels of listener judgments, tens of millions of them, that trained the model, so independent verification is still pending.

Speed and cost are the other two axes where the company claims wins. Hume says the API is fast enough to ask dozens of questions every few seconds on a live call, and to make an archive of 40,000 recordings searchable in minutes, something it argues no existing pipeline can do without transcription, labeling and indexing. It positions generative models and fixed classifiers as the two alternatives, one too slow for structured judgments, the other too rigid to redefine.

The intended buyers are contact centers and voice-agent builders, and Hume is candid about near-term limits: it suggests pairing audio judgments with transcripts before handing both to an LLM, and says live and ambient applications remain a longer-term opportunity as real-time capabilities develop. Access is through a waitlist, with a playground for trying custom questions on uploads or live audio.

The launch fits a wider pattern: as generative voice gets cheap, value migrates to the sensing layer, the part that tells an agent what a human actually meant. Hume's bet is that perception grounded in panels of human listeners, rather than text proxies, is a defensible layer of the voice stack. The 70.7 percent figure will need outside replication, but the interface, questions in and probabilities out, is the genuinely interesting part.