Emotion labeling for speech, tagged per turn on your taxonomy

Words have a ground truth. A transcript is right or it is not, and two good annotators agree almost every time. Affect has no such anchor: one clipped sentence reads as irritation to one listener, fatigue to another, and nothing at all to a third, and none of them is wrong.

That does not make the signal useless, it makes the reporting matter. Every batch is labeled by several annotators and shipped with per-class agreement, so you can weight training on the classes that held up and treat the contested ones with the caution they deserve.

Most teams start with discrete classes per utterance or turn, because the categories map cleanly onto a classifier head. Where a model needs something continuous, valence and arousal ratings are layered onto exactly the same segments, so one pass produces both views of the audio and the two never drift apart.

Intent is handled separately. It can arrive as dialogue acts, or mapped into whatever domain taxonomy you define, because teams almost always want intent expressed in their own categories rather than ours.