Discourse analysis or conversation analysis approaches audio or video data by way of transcription. It takes sentence or utterance as its point of departure, from which it moves up to discourse or conversation, or down to parts of a sentence or utterance. The present study departs from this mainstream paradigm by outlining and demonstrating a corpus linguistic approach to multimodal text analysis that starts from the analytic unit of social situation, to that of activity type, task/episode, and the participants' behavior of talking and doing. The primary data consists of video streams with synchronized sounds rather than orthographic transcripts. The segmentation and annotation of nondiscrete streams of a multimodal text are demonstrated in accordance with the latest Text Encoding Initiative (TEI P4).