A voice agent that answers correctly while the caller gets angrier is still failing. Turn-level affect labels on natural conversation give an agent something to detect, and give you something to evaluate it against.
Most openly available emotion corpora are performed: actors reading set sentences in a named emotion. Models trained on them learn a clean, exaggerated version of affect that real callers rarely produce, and they degrade the moment a genuinely irritated person speaks in a flat voice.
Labeling spontaneous conversation is harder and the agreement figures are lower, which is exactly why we publish them. What you get in return is affect that occurred rather than affect that was staged.
Detecting frustration is only half the requirement. The other half is whether the agent did anything sensible about it, and that cannot be measured without labeled turns to measure against.
The same annotation pass supports both, so an evaluation set and a training set share one schema rather than being reconciled after the fact.
Use the free worksheet to define coverage, rights, delivery fields, held-out rules, and acceptance tests before requesting a collection plan.
Build a dataset specificationName the language and the interaction type. The team labels a representative conversational excerpt with your categories and sends it manually within 48 hours, free of charge.
Enter your work email and verify it with a 6-digit code. The team then selects a representative sample and sends it manually within 48 hours. Use the optional field to name the language, channel, or condition you need to inspect.
Category sets, dimensional scales, and agreement thresholds are usually a conversation, not a form field. Book a short call to work through the taxonomy and conditions the labels need to hold up under.
