Treat labels as acoustic measurements
Scores summarize patterns in speech expression. They are not direct access to a speaker’s private feelings, intentions, honesty, or mental state.
Responsible use
Emotion labels describe expression, not inner truth
oruk measures acoustic and linguistic patterns associated with reported emotion and speaking style. Appropriate applications preserve uncertainty, validate in context, and keep consequential decisions under qualified human control.
Do not claim that the API knows how someone truly feels. The same vocal pattern can have different meanings across people and situations, and a person can deliberately modulate expression.
Keep the score distribution
Do not reduce multilabel output to a single definitive judgment when several labels or styles are plausible.
Validate in context
Expression varies by setting, microphone, language background, culture, health, and individual speaking habits. Evaluate on the population and audio conditions where the system will run.
Keep people in the decision
Use outputs to support review, search, aggregation, and research. Do not use them as the sole basis for employment, insurance, credit, medical, legal, educational, or law-enforcement decisions.
Obtain rights and consent
Tell people when recordings are analyzed and secure all permissions required for collection, processing, storage, and downstream use.
Monitor outcomes
Track subgroup performance, false positives, human overrides, and downstream effects. Pause automation when behavior changes materially.
Recommended product language
Prefer “the speech was classified as sounding frustrated” over “the customer was frustrated.” Show scores, alternatives, time-local segments, and an uncertainty state where appropriate.
Evaluation before launch
Create a representative held-out set, define acceptable errors with domain experts, compare human agreement, evaluate relevant subgroups and recording conditions, and document escalation paths.