(Following a Stradivari trail by Amara Graps.)

Input Data.

The original sound files at the Scientific American website are a pair of 4.5 minute MP3 datasets, of different music for each violin, and performed by the same violinist: Zina Schiff. (The Web site actually lists 3 pairs of comparison datasets but the latter two pairs have a piano accompanying the violinist, so then unusable for this experiment.) I selected from each, the first sixty seconds. Then I mixed the two stereo channels for each violin into one monochannel, and saved each as an ascii file. I selected 50,000 consecutive points (=2sec) at ten different places in each one million point (= 60sec) ascii file.

(The article that began this trail is a lovely blending of history and science and music. I recommend it.)

Violin A	Violin B

Notice that the intensity/loudness for each Bach piece is different (see the y-values). While the difference in loudness might be confusing to our human listening ear to distinguish features of the two violins, I don't think that it matters for the separation of the frequecies in wavelet space. (Please tell me if I'm wrong.)

Violin A is music from J.S. Bach: Sonata No. 1 Adagio

Violin B is music from J.S. Bach: Sonata No. 1 Fugue.

Frequency Transformation.

I converted those ten different data-extractions from each sample into frequency space using a wavelet Daubechies 18 filter and a discrete wavelet transform from Wavelet Workbench.

The lowest frequencies are at the top scales of this two-dimensional plot, and the highest frequencies are located in the bottom scales. One key to answering this question then is: Where is the "activity" located for each violin ??

Violin A	Violin B
Violin A	Violin B
Violin A	Violin B
Violin A	Violin B
Violin A	Violin B
Violin A	Violin B
Violin A	Violin B
Violin A	Violin B
Violin A	Violin B
Violin A	Violin B

Home	Last Modified by Amara Graps on 1 September 2002. © Copyright Amara Graps, 2002.