Compressing Transformer-based self-supervised models for speech processing