A voice recording can carry stories, memories and personal expression. When AI adds an emotion label to that recording, it also adds an interpretation. For anyone exploring automated audio description, an important question follows: how consistently does that interpretation work for different speakers?
Looking beyond average performance
This EUSIPCO 2026 study audits Qwen2-Audio using five English-language speech datasets. The model predicts emotion labels without task-specific training. The researchers examine gender and, where available, age and ethnicity through three fairness measures: prediction balance, recognition of each emotion, and overall accuracy across groups.
Similar predictions can hide unequal reliability
Qwen2-Audio led the tested audio-language models on four of five datasets. Yet similar overall prediction patterns concealed substantial accuracy gaps. On RAVDESS, the reported gender accuracy difference was 46.58 percentage points; on EmoV-DB, it was 36.96.
These figures are dataset-specific. Small groups and uneven emotion distributions increase uncertainty, so the authors caution against comparing fairness values across datasets.
Model settings matter too
Higher generation temperatures, which increase randomness, generally reduced performance. Settings at or below 0.7 offered the best balance in these experiments. The study provides a diagnostic audit, not a bias-removal method; its English-only scope limits generalisation.
What this means for cultural heritage
For REEVALUATE, the practical implication is to treat AI-generated descriptions as information that needs context and review. A museum team exploring audio annotation could ask which recordings were tested, whose voices are represented, and how mistakes will be identified and corrected.
Consider an oral-history recording: a proposed emotion tag might help someone explore it, but should not become an unquestioned statement about the speaker’s feelings. Keeping the original recording accessible and clearly distinguishing machine suggestions from human descriptions would help readers assess the interpretation. This is an illustrative implication for cultural heritage practice, not a museum deployment evaluated in the paper.
Read the publication
Multiclass Fairness Analysis of Qwen2-Audio in Speech Emotion Recognition
Federico D’Asaro; Juan José Márquez Villacís; Andrea Bottino; Giuseppe Rizzo. EUSIPCO 2026.
Read the full open-access paper
REEVALUATE is funded by the European Union under Grant Agreement No. 101132389. Views and opinions expressed are those of the authors and do not necessarily reflect those of the European Union or the European Research Executive Agency. Neither the European Union nor the granting authority can be held responsible for them.
