full-fledged pidgin

Every time I measure speech sounds for some project or to answer some question I find myself wishing that the LSA web site (or some similar organization — ASA?) had a few pages devoted to the nuts and bolts of measuring speech sounds: what trade-offs people make, how to make sure you’re being consistent, etc. Then Keith Johnson, I think jokingly, suggested we, the students in his LSA 07 Experimental Phonology class, should post on our blogs about how we take measurements with illustrative screenshots. I’m pretty sure he was kidding, so I’ve decided to do it.

Disclaimer: I believe that I’m following what Pam Beddor taught me, but any terrible decisions or ridiculous errors are entirely my own.

Measuring a VN pair.

Take for example the measurement of a vowel-nasal pair. My first rule is always to be as consistent as possible. My version of consistent is to be conservative on the left side of a sound and greedy on the right.

Conservative on the left: vowel

Here I’m beginning to measure a vowel’s duration. Naturally, I’ve placed the cursor at a zero crossing, but I’ve done some other things as well. I always start measurements (of periodic sounds) when the sinusoidal curve is beginning to move above the zero crossing. I am also conservative on the left side — starting my measurement at the onset of the first peak that looks like the periodic sound I’m measuring.

Greedy on the right: vowel

Here’s the other end of this vowel. The waveform is returning from a valley, the cursor is at the zero crossing. The waveform no longer looks much like the vowel I’m measuring, but it’s also clearly not the following consonant. Because I was conservative on the left I have to be greedy on the right or I’d get inconsistent results. I could select only the choicest, most beautiful peaks to be in my measurement, but then measurements of, say, a VN pair would either be wildly inconsistent across files or I’d leave out snippets of sound (like that last peak in this selection) that don’t look like clear members of either. One could also be greedy on the left and conservative on the right, but I don’t think you can reasonably implement the same strategy on both sides without making a bit of a mess.

Conservative on the left: nasal

Just to complete the picture, notice that here my cursor begins measuring the nasal consonant at precisely the point I stopped measuring the preceding vowel. I haven’t found an especially easy way to do this in Praat (or any other application). It seems like there should be a “pivot selection around cursor” option or something similar — so I can be absolutely certain I left my cursor in the same spot. I’ve gotten so I can normally do a lot of these measurements pretty quickly, but fussing about this exact point is probably my largest source of remaining delay.

Greedy on the right: nasal

Finally, the nasal consonant measurement extends, in this case, to include every waveform without clear frication (both visible and audible). I don’t claim to be a master at this, I may be doing it entirely wrong, but I’m usually pretty happy with my consistency (when, for example, I repeatedly measure the same 5 files 4 times each and then compare my measurements across trials).