# What is the "pitch marker" in the LP-PSOLA algorithm?

The LP-PSOLAR algorithm uses a the Linear Prediciton Coding (LPC) to calculate an error vector from a signal. This error vector is used to calculate the pitch markers. What do these pitch markers represent in the original signal in generall? Do they represent a change in fundamental frequency or a change in phoneme (I think both assumptions are wrong)? The following figure depicts the error vector (red) and the signal (green). The second picture depicts an additional example.

The pitch marker indicates the beginning of each cycle of the waveform - or from a more physiological point of view the point at each the periodical motion of the glottis causes a sudden change in air flow.

A stable sound (in pitch and phoneme) will still have a marker every cycle - so these markers do not indicate any change.

• I have added a new figure to my post. Am I right that the red points on the second figure represent the pitch marker? Is pitch marker a kind of markter that shows that borders of basic waveform? Sep 13, 2013 at 10:22
• In LP-PSOLA, pitch markers are chosen so that a/ the interval between them correspond to the estimated period of the signal; b/ they optimally coincide with maxima in the excitation signal (residual of LP analysis), which, according to the source-filter model, represents the contribution of the glottis. I don't think your test signal has been generated with a speech production model (harmonic comb sent into a bunch of narrow resonators), so the notion of peak in the excitation signal is a bit ill-defined here... Sep 13, 2013 at 10:56
• I changed the figure to a speech signal. I think I should read something about the human speech production. Sep 13, 2013 at 11:54

Interesting post, I'm working on the same, your plot seems to be catching the valeys Marks, I did the same using Pitch Track based in autocorrelation, but my marks are in Peaks (Maximum excitation of signal).

For my first test I'm using "Kara.segment1.aiff" from Peeter's page, here my plot marks :

When you speech there is a moment where your glottis are closed and other instant where it is open, this moments are find by Pitch Markers, is interesting see how glottis have relation witch Pitch Period, you can get near values of pitch using Pitch Markers.

You can see how near they are, for this plot I get:

Period extracted from Pitch Track (Autocorerlation Based):

395 379 365 359 355 349 342 337 332 329 325 322 319 316 314 312 310 306 302 304 311

Period extracted from Pitch Mark using Peaks:

393 381 368 364 343 348 345 340 327 328 325 323 320 316 317 198 312 426 297 329 299

• My example (the green figure) is calculated with a Linear Prediction Coding. This method uses the Levinson-Durbin recusion. Sep 18, 2013 at 7:36
• great, seems be a nice method, I would like to see how it works, you can share the source with me? Sep 18, 2013 at 11:28
• I'm sorry. I havn't coded it yet. It is described in a thesis, but this thesis is written in german. The tile was: "Algorithmen in Akustik und Computermusik 02" WS/SS 07/08 Thema: (subject) Pitch Shifting & Time Stretching. The website of one of the students is written in english: http://www.benbengler.com/contact.html. Maybe he helps you. I will use another algorithm, based on FFT. It is described here: http://www.katjaas.nl/pitchshift/pitchshift.html. Sep 19, 2013 at 7:37
• Thank you for all informations, some time ago I did it based in FFT using matlab, it's works well ... Sep 19, 2013 at 11:23
• OK. Can an you give me your matlab code, please? ;-) (My e-mail adress is displayed on my profile. Sep 19, 2013 at 17:33