Análise da Robustez das Representações Latentes de Modelos Autossupervisionados para Reconhecimento de Emoções em Sinais de Voz

This work investigated the robustness of Self-Supervised Learning (SSL)-based models for speech emotion recognition, specifically HuBERT and Wav2Vec2, under different acoustic degradation conditions. For this purpose, the IEMOCAP (Interactive Emotional Dyadic Motion Capture Database) dataset was used in combination with synthetic telephone noise generated through MNRU and environmental noises from the MUSAN corpus, added at different intensity levels. The results demonstrated an overall superior performance of the HuBERT model, with greater robustness under the evaluated degraded conditions. For both models, a progressive performance degradation was observed as noise severity increased. HuBERT exhibited high robustness to telephone noise, whereas under babble and crowd noise conditions an increasing tendency of confusion between emotions was observed, particularly toward the neutral class. In contrast, Wav2Vec2 presented a less consistent degradation pattern, with higher sensitivity for emotions such as anger and sadness. The results also suggest that architectural differences between the models directly influence their robustness, with HuBERT's masked prediction mechanism based on discrete units favoring more stable representations under degradation. Furthermore, it was observed that certain types of noise not only degrade the signal but also introduce acoustic characteristics capable of shifting latent representations toward other emotions, such as increased energy and spectral variability approximating some samples to patterns associated with neutrality or happiness.