instrument suit (and test!) many theories of Baroque pitch and other questions related to performance practice. The ability to switch between harpsichord and organ has been particularly valuable in investigating the ´Klavier oder Orgel´ question that remains unresolved for many of Bach's 'keyboard' works.

These experiments with sound control seem to subscribe to what might be called a spreadsheet mentality. The fundamental question is, 'What if...?' What if the user changes the (tempo, dynamics, volume, articulation, etc.)? The attitude is exploratory. The implication is that the information set (i.e., the musical data) is merely an unbiased mass of details that can be shaped to take on many new identities. This idea stands in sharp contrast to the search-engine mentality; yet both are founded on the use of music as information.

New Paradigms: (4) Harmonization as a Rule-Based System

The study of tonal harmony has always been a discipline based on rules especially those of chordal progression and voice leading. In relation to this tradition as it relates to the chorales of Bach, the work of Kemal Ebcioglu, a researcher at the Thomas J. Watson Laboratory in Yorktown Heights, New York, holds a special place. Although a number of researchers have subsequently pursued similar goals by different means, none has, to my knowledge, yet succeeded in going beyond what Ebcioglu has done. His interest was in creating a program to harmonize automatically the chorale tunes set by Bach.

The first stage in his research was to collect various codified rules relating to the harmonization of chorales. From the perspective of musical scholarship, a compellingly interesting part of the this project has been the discovery that this corpus is vastly insufficient to support the harmonization of even such simple melodies.

So Ebcioglu had to analyze a board range of materials in order to greatly expand this list. At last report, his expert system contained more than 300 rules. It had produced almost 200 chorale harmonizations. In a number of cases the same chorale tune was harmonized multiple times with, of course, different results.

The computer, having been presented with a chorale tune, generates a bass. It then generates inner voices that satisfy the needs of the outer voices. At the heart of the activity there is a processing system that relies heavily on a back-propagation algorithm to inspect the resulting harmonies and voice-leading from multiple viewpoints (see Ex. 3).

Although organists will readily detect the shortcomings of automatic harmonization if they are able to see a score, listeners have some difficulty in making this discrimination when no score is available.

A series of experiments with taped performances of the chorale O Haupt voll Blut und Wunden was made more complicated by performing harmonizations by Bach and