UMI is a good thing. If you ask them nicely they will send you information about and links to all the newly uploaded doctoral dissertations in North America. Today I came across an interesting thesis from McGill University by Ian Knopke (reference below). Ian worked on a music search engine that, unlike the standard search engines – not only analyzes the textual information of web pages but also the metadata from the sound files and the audio files themselves! What kind of audio file data is Ian’s system using? You have to read his dissertation up to page 126 to find an answer: Ian’s web crawler uses a software kit named MARSYAS. This tool can perform essentially two analyses on music files. First, it can classify the genre of the file. Second, it can classify whether a file contains music or spoken text. So, this basically means that the audio file analysis part of Ian’s search engine can tell you if a file actually is music and where to find it in your local record store (“Heavy Metal”, “Classical”, “Country”).

One wonders whether this type of information will dramatically improve search results. Can’t this kind of information be extracted from the textual information on the website that refers to the audio file already? Apparently not, but Ian does not seem to be too happy about the results of the audio file analysis either, for in the appendix of his dissertation he proposes to enhance the external and metadata levels by including a new AUDIO tag into the html specifications. “If information such as sampling rates or genre were accessible from web pages alone, it would probably not be necessary in many cases to retrieve the audio files. Appendix A proposes a new AUDIO HTML tag that, if implemented, would overcome many of these limitations.” (p. 153) Ironically, this new AUDIO HTML tag would result in a situation in which an audio search engine that analyzes sound files will become obsolete.

Music analysis of sound files is a field in Artifical Intelligence that still needs a lot of basic research before it will produce satisfactory results that can be used in applications. Once we can extract data on instrumentation, tempo, key as well as more detailed information such as articulation and timbre, Ian’s search engine will be a fantastic tool. Until then we will have to rely on our human dealer in the local record store. It’s not the worst choice anyhow.

[Knopke, I. 2005. Building a Search Engine for Music and Audio on the World Wide Web. Ph.D. Thesis. McGill University, Canada.]