Speech-based emotion recognition with self-supervised models using attentive channel-wise correlations and label smoothing