MooseNet: A Trainable Metric for Synthesized Speech with a PLDA Module