In this work we propose a versatile approach to an automatic video summarization problem. In contrast to rigid schemes it allows a user to specify his preferences concerning desirable content of a summary by imposing constraints on the video segments features. Multiple signal-level and semantic-level features extracted from both the audio and the video streams are combined in this way in order to select relevant video fragments for a summary. As these fragments are selected globally for the whole video, some higher level semantic segments may be missed in the summary. To avoid this, amodified technique is proposed that generates a kind of a summary which we call a digest that preserves the semantic structure of the video.