Professor Leslie S. Smith: Statistics of Natural Sound: some PCA's and ICAs

Statistics of Natural Sound: Some PCAs and ICAs

The PCAs and ICAs below were produced using the ridiculously simple technique of using the digital sound directly as the input vectors. The sounds were (randomly) cut up into 200ms long sections. This technique has the advantage of being invertible in the sense that we can play back the PCAs and ICAs by considering the projection mapping as a sound. (I also use other techniques for examining PCAs and ICAs: see my seminar (IMNS, Edinburgh 4 May 2004).

The PCAs and ICAs were produced using MATLAB running under Windows XP. This has a limitation of about 1.3Gbytes for the programme when it is running, (and then it runs out of memory). This has restricted the size of dataset I could investigate. Of course, that's me blaming the tools, it is certainly true that had I time, I could rewrite the code (in C++ perhaps) so that it did not have this problem - or I could have used a different approach to PCA and ICA estimation that did not involve matrix inversion. Nonetheless, that is the limitation I have been working within.

PCAs and ICAs from a mixed set of TIMIT utterances.

The utterances used were a mixed set of 40 TIMIT utterances. The PCAs, ICAs are all .au files. They are all 200ms long. Sampling rate 16000Samples/second, 16 bits linear.

PCA 1-10	PCA1	PCA2	PCA3	PCA4	PCA5	PCA6	PCA7	PCA8	PCA9	PCA10
PCA 11-20	PCA11	PCA12	PCA13	PCA14	PCA15	PCA16	PCA17	PCA18	PCA19	PCA20
ICA 1-10	ICA1	ICA2	ICA3	ICA4	ICA5	ICA6	ICA7	ICA8	ICA9	ICA10
ICA 11-20	ICA11	ICA12	ICA13	ICA14	ICA15	ICA16	ICA17	ICA18	ICA19	ICA20

PCAs and ICAs from the first minute or so of Coltrane's Equinox (a Jazz piece).

This time the sampling rate is 22050 samples/second.

PCA 1-10	PCA1	PCA2	PCA3	PCA4	PCA5	PCA6	PCA7	PCA8	PCA9	PCA10
PCA 11-20	PCA11	PCA12	PCA13	PCA14	PCA15	PCA16	PCA17	PCA18	PCA19	PCA20
ICA 1-10	ICA1	ICA2	ICA3	ICA4	ICA5	ICA6	ICA7	ICA8	ICA9	ICA10
ICA 11-20	ICA11	ICA12	ICA13	ICA14	ICA15	ICA16	ICA17	ICA18	ICA19	ICA20

Back to Professor Smith's page on statistics of natural sound.

Last updated: Thursday, 06-May-2004 10:02:46 BST

If you have any difficulties accessing this page, or you have any queries/suggestions arising from this page, please email:
Prof Leslie S Smith (lss(nospam_please)@cs.stir.ac.uk)