Date of Award


Document Type

Thesis (Ph.D.)


Department of Computer Science


This thesis introduces multiple media correlation, a new technology for the automatic alignment of multiple media objects such as text, audio, and video. This research began with the question: what can be learned when multiple multimedia components are analyzed simultaneously? Most ongoing research in computational multimedia has focused on queries, indexing, and retrieval within a single media type. Video is compressed and searched independently of audio, text is indexed without regard to temporal relationships it may have to other media data. Multiple media correlation provides a framework for locating and exploiting correlations between multiple, potentially heterogeneous, media streams. The goal is computed synchronization, the determination of temporal and spatial alignments that optimize a correlation function and indicate commonality and synchronization between media objects. The model also provides a basis for comparison of media in unrelated domains. There are many real-world applications for this technology, including speaker localization, musical score alignment, and degraded media realignment. Two applications, text-to-speech alignment and parallel text alignment, are described in detail with experimental validation. Text-to-speech alignment computes the alignment between a textual transcript and speech-based audio. The presented solutions are effective for a wide variety of content and are useful not only for retrieval of content, but in support of automatic captioning of movies and video. Parallel text alignment provides a tool for the comparison of alternative translations of the same document that is particularly useful to the classics scholar interested in comparing translation techniques or styles. The results presented in this thesis include (a) new media models more useful in analysis applications, (b) a theoretical model for multiple media correlation, (c) two practical application solutions that have wide-spread applicability, and (d) Xtrieve, a multimedia database retrieval system that demonstrates this new technology and demonstrates application of multiple media correlation to information retrieval. This thesis demonstrates that computed alignment of media objects is practical and can provide immediate solutions to many information retrieval and content presentation problems. It also introduces a new area for research in media data analysis.


Originally posted in the Dartmouth College Computer Science Technical Report Series, number PCS-TR98-335.