This paper investigates sentence compression for automatic subtitle generation using supervised machine learning. We present a method for sentence compression as well as discuss generation of training data from compressed Finnish sentences, and different approaches to the problem. The method we present outperforms state-of-the-art baseline in both automatic and human  valuation. On real data, 44.9% of the sentences produced by the compression algorithm have been judged to be useable as-is or after minor edits.