Unsupervised Pre-Training of Bidirectional Speech Encoders via Masked Reconstruction