Reproducibility beyond the Laboratory

This past Sunday I was writing about the problem of seeking reproducible results when studying the making of music, which, of necessity, is "entangled" in history. These thoughts have been nagging me again while I have been reading and old (1982) review paper, "Perception of Singing" by Johann Sundberg, which was included in Diana Deutsch's The Psychology of Music Collection. I have been reading about Sundberg's efforts in the more recent context of a much more recent practice called "auditory scene analysis," which was in its infancy (if it existed at all) back in 1982.

Auditory scene analysis amounts to the ability to identify and extract the different sound sources when given a "real-world" recording. I remember seeing a demonstration based on a studio recording that had been disrupted because someone dropped a bottle while the musicians were performing. Through auditory scene analysis, it was (at least in theory) possible to isolate that source of "noise" and remove it, leaving only the sounds of the performing musicians.

It has been over a decade since I followed research in this field, so I have no idea how far it has progressed. However, it occurs to me that such a discipline would provide tools by which we could spend more time studying recordings made "in the field" (which could be in commercial recording studios or audio capture of "live" performances) and less time worrying about creating artificial situations in laboratory settings. I was reminded of a talk I once heard by a Google researcher who claimed that one could make more progress in natural language understanding by the statistical study of the vast number of documents that now exist in digital form on the Internet (and have been indexed by Google) than through traditional practices of linguistic research.

While I was not entirely convinced, it seems to me that, where music is concerned, we now have a prodigious amount of recorded "field data," whether or not we put all those data to use. If our intelligence agencies are now using large quantities of recorded telephone calls to create and test hypotheses, why can't we develop methods for doing the same sort of thing with recordings of musical performances? If nothing else the consequences of a failed hypothesis are likely to be less crucial (and enough of those recordings have been made generally available that we would not have to worry about invading anyone's privacy).