Existence of per-word emotion-annotated speech data

Determine whether annotated training data currently exists in which emotions are labeled separately for each word, rather than for entire sentences or passages, to support per-word emotion tagging in Deaf-centric text-to-speech systems.

Background

The paper identifies per-word emotion tagging as a participant-driven requirement for Deaf-centric text-to-speech interfaces. Implementing this capability would require training data that associates emotional labels with individual words rather than assigning a single emotional annotation to an entire sentence or passage. The availability of such data is explicitly unresolved, and its absence would constitute a major technical barrier to fine-grained, non-auditory control of emotional speech output.

References

Per-word, as opposed to per-sentence, emotion tagging requires annotated training data that has emotions annotated separately for each word, rather than entire sentences or passages. It is unclear whether such data currently exists.

— Seeing the Voice, Preserving the Self: A Participatory Design Approach to Deaf-Centric Text-to-Speech  (2609.10199 - Atemnkeng et al., 9 Sep 2026) in Section 5, “Technical Requirements for Implementing the Designs,” subsection “Emotion Customization”