General-Purpose Speech Representation Learning through a Self-Supervised Multi-Granularity Framework.