Use Case Scenario: Find a (given) 'core' passage of music known to recur at least once in each of a collection of recordings of the Bach/Gounod 'Ave Maria'.

Short Description

Gounod's song 'Ave Maria' and the instrumental piece that preceded its composition are based on JS Bach's first prelude in C from the 48 Preludes & Fugues. This was incredibly popular in the late 19th-early 20th centuries, around the time recordings began. As well as at least 20 editions of the sheet music, the British Library holds (in the British Sound Archive) over 600 recordings from c.1902 almost to the present day.

The Bach prelude will be present in all recordings in some version that is more or less close to the score Gounod published - although in some arrangements it is all but inaudible, in which case only the harmonic outline remains (though even this is distorted in some extreme examples).

One is reminded of the smile on the face of Lewis Carroll's Cheshire Cat - see http://en.wikipedia.org/wiki/Cheshire_Cat

Beneficiaries

Musicologists, but also others (?)

Task Description using fftExtract and AudioDB

Can be done for some cases out of pilot test set of 37 recordings using chroma, but a more explicitly 'harmonic' feature might work better. NB Noise on historical recordings is definitely a problem.

Requires intensive processing of audioDB result-output over multiple searches. (This may be a general characteristic of musicological use cases.)

Data Set and data set preparation

Ground Truth

I have a first test-set of 37 recordings (NB not rights-free); while all can be assumed to contain vestiges of the Bach, this is not all that easy to 'hear' in many cases.

The musical 'core' (ie the Bach) could be presented either as a solo (piano or harpsichord) performance of the original Bach prelude, as Gounod's own 'Meditation sur la premiere prelude de J.S. Bach' for violin/cello and piano (his first version), or as an extract from one of the Ave Maria recordings themselves.

Fairly precise location (timing) data for the 'core' in each recording could be obtained by listening and annotation (eg in Audacity or Sonic Visualiser), but this has not yet been done.

First tests were done by listening to likely passages identified by audioDB; they did not properly account for false positives which were simply ignored (!!).

From audioDB 'best' results I could determine basic pitch-standard of the performance (using 36-d chroma) and the 'variability' of pitch (e.g. resulting from vibrato or general poor intonation). Crude timing data was used to derive average relative tempos. Presented at CIM08 (Thessaloniki, July 08).

Experimental Method

Results Tables and Figures

Conclusions

-- TimCrawford - 17 Sep 2008