Hume's Audio Decision API answers your own questions about a recording, like "is the caller frustrated?"

Hume launched its Audio Decision API on November 2. The idea: instead of transcribing a call and running text models over the words, you ask the audio directly. You write each question as a sentence, list the possible answers, and the model returns a probability for each one.

There are four question types. In the API they're called noul (yes or no), choice, score and span. Hume's examples include "Does the customer sound frustrated?", "How many speakers can you hear?" and "Where does someone laugh, sigh or gasp?", which returns start and end times. You can ask several questions about the same file in one call and add context to the request.

The endpoint is POST /v0/decisions/file. A second endpoint, POST /v0/decisions/file/input_tokens, estimates the input tokens of a request before you run it, and run endpoints return the status, timing and audio size of past requests. There are Python (hume-audio-decisions) and TypeScript (@humeai/audio-decisions) SDKs.

Hume says that in an initial evaluation the model agreed with a human panel's average judgment 70.7% of the time, against 56.5% for a single human rater. FourWeekMBA notes Hume didn't say how many recordings or raters were used, or which languages. Access is through a request form for now, with a playground for live and uploaded audio.