Demonstrates that a speech-emotion-recognition model which is gender-fair on its source corpus does not stay fair after cross-corpus transfer, because source-only debiasing (Fairway, Reweigh) and standard transfer-learning methods fail to generalize fairness to the target corpus. Proposes Combined Fairness Adaptation (CFA): a wav2vec-2.0/transformer emotion classifier with an auxiliary adversarial gender classifier connected through a gradient-reversal layer that strips gender information from the emotion features of both source and target corpora, jointly optimizing cross-corpus recognition and gender neutrality and significantly reducing demographic-parity and equalized-odds gaps on the target corpus.
This machine-learning paper investigates whether the gender fairness of a speech emotion recognition (SER) model survives cross-corpus transfer. Cross-corpus SER trains on one speech corpus and is de…