Hume Expression Measurement Api

Operator submits a batch job over a set of audio files and wants both vocal-prosody and spoken-language expression signals.

In the POST /v0/batch/jobs models config, enable the modalities that match the media and goal (prosody for vocal expression, language for transcript-derived expression); only enable face for video/image inputs. Each model returns its own per-frame/per-utterance expression scores.

Audio contains laughs, sighs, and gasps the operator wants characterized separately from spoken prosody.

Enable the vocal-burst model for non-linguistic vocalizations; keep its outputs distinct from the prosody model's speech-based scores. Do not merge burst and prosody outputs into one undifferentiated score set.

A prediction frame shows the 'anger' expression score at a high value for a customer-support call. A downstream system wants to flag the customer as 'angry'.

Treat the score as the intensity of a perceived expression, not a verified internal emotional state. Surface it as a probabilistic signal with appropriate hedging ('voice shows expressions associated with anger'), never as a definitive classification of how the person actually feels.