ERM-MinMaxGAP: Benchmarking and Mitigating Gender Bias in Multilingual Multimodal Speech-LLM Emotion Recognition

Speech Emotion Recognition (SER) systems have growing applications in sensitive domains such as mental health and education, where biased predictions can cause harm. Traditional fairness metrics, such as Equalised Odds and Demographic Parity, often overlook the joint dependency between demographic attributes and model predictions. We propose a fairness modelling approach for SER that explicitly captures allocative bias by learning the joint relationship between demographic attributes and model error. We validate our fairness metric on synthetic data, then apply it to evaluate HuBERT and WavLM models finetuned on the CREMA-D dataset. Our results indicate that the proposed fairness model captures more mutual information between protected attributes and biases and quantifies the absolute contribution of individual attributes to bias in SSL-based SER models. Additionally, our analysis reveals indications of gender bias in both HuBERT and WavLM.

Current state-of-the-art Speech Emotion Recognition (SER) systems rely on lots of training data, which is available only for a few languages such as English. For low-resource languages, using a cross-lingual model or adapting it to the target language might be an alternative to training a mono-lingual model from scratch. However, it is not clear what constitutes a "low-resource", i.e. where model adaptation loses its edge over the mono-lingual approach. We investigate this topic by training and adapting several SER models on English and Mandarin data, focusing on the emotional attributes of arousal, valence and dominance. Our results show that, for arousal, model adaptation is beneficial even with two hours of data, while for valence 10-20 hours of adaptation material is required to match the performance of a mono-lingual model trained with 100 hours of data. This indicates that task-specific model adaptation may be an effective approach in low-resource speech emotion recognition problems.

As large neural models have become better at language tasks, researchers are increasingly building multi- and omnimodal models that handle more modalities of data. One example is the expansion of speech recognition models to audio-visual data for noise mitigation and multimodal subtitling. While performance and bias have been studied extensively in the single-modality regime, it is unknown how new modalities affect this, even though they produce biases in humans. We therefore propose the first bias evaluation of multimodal speech recognition, where we create videos pairing different faces with the same audio, and measure changes in speech transcription accuracy. We find large quality-of-service differences across mWhisper-Flamingo and Gemini models, with drops of up to 4.05 word error rate points, across self-declared gender, ethnicity, and their intersection. Our findings point to a priority for developers to evaluate, fix, and communicate such limitations, as providing more signals through additional modalities is not necessarily better, and may even lead to biased outcomes.