Learning contextual relevance of audio segments using discriminative models over AUD sequences

摘要

Effective retrieval of multimodal data involves performing accurate segmentation and analysis of such data. With easy access to a number of audio and video sharing platforms online, user-generated content with considerably less than ideal recording conditions has increased rapidly. One major issue with such content is the presence of semantically irrelevant segments in such recordings. This leads to the presence of considerable contextual noise in such recordings that makes analysis difficult. In this paper, we present a discriminative large-margin based approach that uses annotated data to understand which parts of the audio are relevant (while noting that the notion of relevance could be extremely subjective and potentially challenging to define), and can automatically extract such segments from new audio.