Look Who's Talking: Inferring Speaker Attributes from Personal Longitudinal Dialog