I have been working with a colleague, Wei Kunlin at Peking University, on the distribution of note durations in classical music, and we have been trying to use MIDI data available on the web as a source. Our interest has been to find the durations of the longest notes in a range of musical performances, to see if there is some absolute limit on note duration independent of performance tempo. But this has proven more difficult than one might think, as there are a number non-trivial problems in the use of this data.
The first is source identification. One can find a file labelled "Mozart Piano Sonata," but then you have no other meta-data: which sonata is it? What is its original tempo marking, key, movement/form, and so forth (all of which may have a bearing on the performed durations). And who is the performer? Expert? Novice? Are the irregularities in timing due to inept performance or judiciously applied rubato?
The second is that multipart musical textures are often all put into a single (or relatively small) number of MIDI channels--so one has to separate out the voices to capture the underlying note durations. Even where different parts have been put in separate channels this can arise, as a pianist's hand can play more than one note, a violin part may have double stops, etc., etc., as Pawel notes below. If one just looks at the interval from note onset to note onset--which would seem sensible--there are very small differences in onset timings for notes that are ostensibly simultaneous; Caroline Palmer has studied these "melody leads" in piano performance and found they are used to convey structural information. So a lot of careful data cleanup has to be done, and much of it has to be done by hand.
Finally there are differences between apparent and effective duration(s) of notes due to auditory streaming effects, as recently discussed by Robert Gjerdingen on this list (the "Perceptual basis of evolving western musical styles" thread)--what musicians call "compound melody" can create longer effective durations as a melody skips through different registers.
Nonetheless, if one can address these problems, there is a rich supply of real performance data out there . . .
P.S. acutally, not many musicologists have done this kind of analysis using real audio or MIDI data, though there have been studies of duration that are score-based. I believe this is because musicologists--especially music theorists--don't often think of approaching problems this way (i.e., using empirical methods).
On Jun 13, 2013, at 8:59 AM, PaweÅ KuÅmierek wrote: