Pre-trained Model Representations and their Robustness against Noise for Speech Emotion Analysis