Federico Flego

Pitch Estimation

To test several pitch extraction algorithms a common used database is, among others, the Keele Database.

Part of my past research activities was devoted to re-estimate the keele pitch labels basing on the original laryngograph (lx) data. I refer to the new pitch labels as the "corrected" pitch values.

The reason for using the term "corrected" in inverted commas, stems from a very interesting discussion I had with Dr. George Meyer, from Keele University.

When I asked Dr. Meyers his opinion on the advantages of having pitch reference estimate with no doubling/halving errors, he briefly, pointed out that:

"There is *no* correct F0 estimate because it depends on the analysis window size and type (among other things) so that we decided that it would be wrong to give a 'correct' reference."

While my belief is that:

"A pitch extractor algorithm, disregarding the analysis frame size, could be so smart to avoid doubling/halving errors, or at least try to do it (maybe also at run-time basing on past-history). In this case, having a correct reference pitch signal should provide, by my point of view, the best feed-back on the quality of the tested algorithm."

I let you draw your own conclusions on the matter but please, remember that you cannot compare the results obtained from a pitch estimation algorithm using these "corrected" pitch estimates, with the results available from the literature where the original Keele pitch estimates were used.

The following shows the differences between the Keele orginal pitch reference values and the "corrected" one. If you can't display the image or it has too poor resolution, click here!

Top axes show the speech signal relative to the file: m1nw0000.pes

Second axes show the pitch estimates in samples. The different plots correspond, as shown in the legend, to:

keele	->	the original keele pitch (red line)
wautoclxHP	->	pitch computed by means of wautoc applied to the filtered lx signal
wautoclx	->	pitch computed by means of wautoc applied to the lx signal
praatlxHP	->	pitch computed by praat on the high-pass filtered lx signal and checked by me (black stars)
praatSIG	->	pitch computed by praat on the audio signal (.pes)

Third axes shows the original Keele laryngograph signal file: m1nw0000.pel
Bottom axes shows the Keele laryngograph signal after high-pass filtering in order to remove the DC component.

In the specific example above it is evident how the original reference pitch estimate (red line) jumps from 0 to the "correct" pitch value (black stars).

In case you wish to examine other differences between the original Keele pitch estimates and those I obtained, relative to the m1nw0000.pes file, here (10 Mb!) there's the matlab .fig file which allows to zoom in every voiced section.

If you're interested in the "corrected" pitch estimates relative to the whole Keele database, saved the Praat format, download the following: KeeleCorrectedPitch.tar.gz.