This document describes a method for automatically extracting key terms from spoken documents. It uses branching entropy to identify phrases, then extracts prosodic, lexical, and semantic features for machine learning. Three learning methods - K-means, AdaBoost, and neural networks - are evaluated. The best performance is from neural networks using all feature types. When applied to lecture transcripts, it achieves an F-measure of 67.31% for key terms, only slightly lower than human annotations.