In this work, we present an approach for semantically based search for similar segments of a speech task, i.e., the search for audio segments in a row audio repository that are semantically related to the audio segment given by a user. Our approach is based on the lexical representation of segments of words that are enriched by semantic relations. We have studied different distance measures and the lexical/semantic representation of the segments. We present experiments for a task of recorded dialogs between students talking about whatever they want, which is a semantically unbounded task. The results, which are encouraging, indicate the potential advantages of using this approach to address this problem.